International Publications – CVAR Lab

2025 - International journal
(7)

Bagging Detection Using Transformer and Multiple Feature Fusion
Nguyen Dinh Vinh, Nguyen Hoang Kha
Nguyen Dinh VinhCorresponding
Nguyen Hoang Kha
Springer Nature Link
Published: February 20, 2025

Object detection involves identifying and localizing objects within images or videos, combining classification and localization. Its importance has surged recently, driven by its broad range of practical uses. This research develops a new deep learning model approach for bagging detection and localization in conveyor systems for industrial applications. The proposed method was introduced by investigating multiple branches along with the proposed general derivative pattern for feature fusion. Experimental results show that the proposed method achieved performance comparable to the state-of-the-art object detection systems under normal input conditions. Moreover, the proposed system obtained better results than existing object detection methods under various testing conditions

deep learning
object detection
performance comparison
localization
classification
conveyor systems
industrial applications
feature fusion
multiple branches
general derivative pattern
state-of-the-art methods
testing conditions
bagging detection
image analysis
video analysis
Blueberry Leaf Disease Classification Using Vision Transformer Technique and Weighted Feature Fusion
Nguyen Dinh Vinh, Narayan C. Debnath, Ngoc Ngo Phuong
Nguyen Dinh VinhCorresponding
Narayan C. Debnath
Ngoc Ngo Phuong
Springer Nature Link
Published: January 1, 2025

Many plant diseases manifest visible symptoms, typically diagnosed by experienced plant pathologists who visually inspect infected plant leaves. However, this manual diagnostic process is slow and heavily dependent on the pathologist's expertise, highlighting its suitability for computer-aided diagnostic systems. Unlike traditional machine learning approaches that demand meticulous manual feature extraction, there is a demand for models capable of successful classification without extensive preprocessing. In this research, we introduced an efficient module that leverages the advantages of vision transformer and weighted feature fusion across raw images, enhanced images, and edge information. Experimental results demonstrated that our proposed system outperformed existing methods such as EfficientNet or DenseNet. Specifically, our proposed system achieved a detection rate and false alarm rate of 98.18% and 0.88%, respectively. In contrast, the EfficientNet attained a detection rate and false alarm rate of 94.24% and 1.56%, respectively, while the DenseNet reached 95.13% and 1.25% for detection rate and false alarm rate, respectively.

deep learning
efficientnet
classification
image analysis
plant disease detection
computer-aided diagnosis
vision transformer
weighted feature fusion
raw images
enhanced images
edge information
manual feature extraction
densenet
detection rate
false alarm rate
plant pathology
Leaf Disease Detection in Blueberry Using Efficient Semi-supervised Learning Approach
Nguyen Dinh Vinh, Ngoc Ngo Phuong, Narayan C. Debnath
Nguyen Dinh VinhCorresponding
Ngoc Ngo Phuong
Narayan C. Debnath
Springer Nature Link
Published: March 1, 2025

Blueberry leaf disease detection is really important to help farmers to early detect leaf disease and find a suitable method to cure the disease. Therefore, this research introduces an approach to detect and classify blueberry leaf disease by using an unsupervised method (auto-encoder), and a supervised method (support vector machine). The accuracy of our proposed method was evaluated by conducting the experiments on blueberry dataset captured at Can Tho City, Vietnam. The existing augmentation techniques was also applied to increase the data size of training and testing. For the first experiment on normal capturing conditions, the F1 scores of the proposed method and SVM are 89.28% and 81.48%, respectively. For the second experiment with noisy conditions, the F1 scores of the proposed method and SVM are 81.5% and 66.7%, respectively.

computer vision
svm
blueberry leaf disease
disease detection
early diagnosis
auto-encoder
support vector machine
unsupervised learning
supervised learning
image classification
can tho city
vietnam
data augmentation
noisy conditions
f1 score
agricultural technology
plant health
leaf image analysis
Leveraging Social Media Data to Analyze Public Sentiment on Mekong Delta’s Environmental Issues
Nguyen Dinh Vinh, Dang Thai Doan
Nguyen Dinh VinhCorresponding
Dang Thai Doan
Springer Nature Link
Published: February 20, 2025

This study leverages social media data from X.com and advanced deep learning techniques to analyze public perception of climate and environmental issues in the Mekong Delta. A dataset of 3,009 tweets collected in May 2024 was analyzed using BERTopic to identify key themes, revealing prominent topics related to water governance, climate change resilience, biodiversity, flood management, agriculture, tourism, and hydropower impacts. These topics were grouped into three themes: (1) hydropower impact and regional tensions; (2) water governance and environmental resilience; and (3) climate change and agricultural impacts. Subsequently, BERT was used for sentiment analysis to classify the sentiments within these topics into positive, negative, and neutral categories. The sentiment analysis indicated that negative sentiment is higher compared to positive and neutral sentiments, with an increasing trend over time. The findings contribute to the academic understanding of social media’s role in environmental discourse and offer practical implications for policymakers and stakeholders in the Mekong Delta. This research underscores the potential of leveraging digital platforms and advanced analytical methods to gain valuable insights for effective climate adaptation and environmental management strategies.

deep learning
social media analysis
x.com
public perception
climate issues
environmental issues
mekong delta
bertopic
water governance
climate change resilience
biodiversity
flood management
agriculture
tourism
hydropower impact
regional tensions
sentiment analysis
bert
positive sentiment
negative sentiment
neutral sentiment
environmental discourse
policymakers
stakeholders
climate adaptation
environmental management
digital platforms
Robust Adaptive Masked Face Recognition Using Mediapipe and Advanced ResNet50 with Multi-Layer Feature Fusion
Le Huy Hoang, Nguyen Thi My Ai, Nguyen Dinh Vinh
Le Huy HoangCorresponding
Nguyen Thi My Ai
Nguyen Dinh Vinh
The 25th International Conference on Computational Science and Its Applications
Published: May 22, 2025

Since the COVID-19 pandemic, wearing masks has become a common practice worldwide, further compounded by the growing issue of air pollution. As masks have become an essential part of daily life, they pose a significant challenge to facial recognition systems, which struggle to accurately identify individuals when key facial features are obscured. This study focuses on improving the performance of existing masked facial recognition systems to better adapt to these real-world conditions. This work builds on an existing framework for masked face detection, proposing several optimizations to enhance recognition accuracy. The approach integrates Mediapipe for reliable face detection, combined with ResNet50 for feature extraction and a novel Feature Concatenation technique that merges multiple facial feature vectors to create more robust representations. To improve model adaptability to masked faces, we employ data augmentation techniques that simulate various real-world conditions, such as different mask types and lighting variations. Additionally, we performed ResNet50 fine-tuning on a specific masked dataset to increase robustness against partial face occlusions. Recognition is then carried out using a cosine similarity-based distance metric, ensuring both computational efficiency and high accuracy. Our optimizations resulted in a significant performance improvement, achieving a precision of 99.47\% on the Celebrity dataset. These results demonstrate the potential to refine existing methodologies for better masked facial recognition, contributing to more reliable applications in security, healthcare, and other sectors.

finetune
restnet50
face recognition
Robust Student Attendance Checking System Using Efficient LBP-based Ensemble Learning Approaches
Nguyen Thi My Ai, Le Huy Hoang, Nguyen Huy Khanh, Huynh Hung Loi, Tran Dai Loi, Nguyen Dinh Vinh
Nguyen Thi My AiCorresponding
Le Huy Hoang
Nguyen Huy Khanh
Huynh Hung Loi
Tran Dai Loi
Nguyen Dinh Vinh
11th International Conference on Intelligent Information Technology
Published: May 14, 2025

Facial recognition systems often face challenges in balancing ac- curacy and computational efficiency, particularly in real-world en- vironments with varying lighting conditions and image quality. Many existing methods rely on individual machine learning mod- els, which may not fully capture the complex patterns needed for high-precision recognition across diverse datasets. This research addresses these limitations by implementing and comparing several machine learning algorithms—K-Nearest Neighbors (KNN), Ran- dom Forest, and Support Vector Machine (SVM)—and improving their performance through ensemble techniques such as Voting and Stacking. The process includes key steps like image preprocessing, facial detection using "hog" and "cnn" models, and feature extraction with facial encodings and Local Binary Pattern (LBP). Additionally, extensive data augmentation techniques—including image rotation, noise injection, LBP, and brightness adjustment—were applied to simulate real-world variations and improve model robustness. Di- mensionality reduction with principal component analysis(PCA) and hyperparameter tuning via GridSearchCV further optimized performance. Our proposed method achieved a facial recognition accuracy of 99.22%, exceeding the results of the SVM 95.34% [15]. This demonstrates the effectiveness of integrating machine learning models and applying ensemble methods, leading to more reliable facial recognition systems and marking a significant improvement over current approaches.

face recognition
recognition
Traffic Sign Classification Using an Ensemble Technique with Machine Learning and Deep Learning Methods
Nguyen Dinh Vinh, Narayan C. Debnath
Nguyen Dinh VinhCorresponding
Narayan C. Debnath
Springer Nature Link
Published: February 21, 2025

Traffic sign classification remains a significant challenge for autonomous vehicles, particularly in recognizing targets of various scales and achieving real-time performance. The varying sizes of traffic signs can negatively affect detection accuracy. Additionally, changes in lighting when capturing input images can pose challenges for existing methods. Therefore, this research presents a new approach to enhance the performance of current traffic sign classification systems by combining the benefits of existing machine learning-based and deep learning-based methods through ensemble learning techniques. Experimental results show that the proposed method improved the performance of existing methods using the German Traffic Sign Recognition Benchmark dataset. Under normal input conditions, the proposed method achieved an accuracy of 95.77%, comparable to the state-of-the-art method YOLOv8, which had an accuracy of 94.97%. However, under difficult input conditions, the proposed method significantly improved YOLOv8's performance by approximately 7%.

deep learning
machine learning
traffic sign classification
autonomous vehicles
real-time performance
varying scales
lighting conditions
ensemble learning
german traffic sign recognition benchmark
yolov8
accuracy
difficult input conditions
performance improvement
classification systems

2025 - International conferences
(2)

Canine Eye Disease Diagnosis: A Hybrid Approach Using YOLOv11 and Autoencoder Models
Nguyen Thi Diem Huong, Ho Van Loc An, Mai Nguyen Truc Vy, Nguyen Vu Duc, Vo Minh Thong, Nguyen Dinh Vinh
Nguyen Thi Diem HuongCorresponding
Ho Van Loc An
Mai Nguyen Truc Vy
Nguyen Vu Duc
Vo Minh Thong
Nguyen Dinh Vinh
International Conference on Intelligent Information Technology
Published: February 20, 2025

The rising prevalence of eye diseases in dogs underscores the need for prompt and precise diagnostic tools to ensure timely treatment and improved outcomes. Traditional diagnostic methods often require specialized veterinary expertise and are prone to delays in detection. This study presents a novel image-based diagnostic framework utilizing a hybrid approach that combines the YOLOv11 object detection model with an autoencoder for feature extraction and disease classification. By employing the autoencoder to distill critical features and leveraging its synergy with YOLOv11 for localization, the system achieves exceptional performance in diagnosing canine eye diseases. Our approach achieves a mAP@50 of 99.5\% and a mAP@50-95 of 90.9\%, demonstrating its robust accuracy and effectiveness. This innovative diagnostic tool offers significant potential to assist veterinary professionals and pet owners in the rapid and accurate identification of canine eye diseases, setting a new benchmark in veterinary diagnostics.

deep learning
data augmentation
autoencoder
pet disease diagnosis
eye diseases
yolov11
Multimodal Approach for Canine Dermatological and Ophthalmological Disease Diagnosis Using YOLOv11 with Data Augmentation and Autoencoder Techniques
Nguyen Thi Diem Huong, Ho Van Loc An, Nguyen Dinh Vinh
Nguyen Thi Diem HuongCorresponding
Ho Van Loc An
Nguyen Dinh Vinh
International Conference on Computational Science and Its Applications
Published: June 30, 2025

Accurate detection of canine dermatological and ophthalmological diseases is essential for improving animal health and supporting veterinary diagnostics. This study introduces a comprehensive approach that integrates the YOLOv11 object detection model with advanced data augmentation and feature extraction techniques, including Autoencoder (AE), CutMix, and Local Binary Patterns (LBP). YOLOv11 provides precise localization and classification of lesions, while Autoencoders enhance feature representation by reducing noise and improving generalization. CutMix increases data diversity, aiding in model robustness, and LBP captures critical texture details relevant to skin and eye conditions. Twelve datasets were created, combining these techniques in various configurations to assess their individual and synergistic effects on model performance. The proposed method achieves a mean Average Precision at 50\% Intersection over Union (mAP@50) of 97.2\% on the original image dataset and 86.6\% on complex images when using a hybrid dataset that integrates original images with LBP, CutMix, and Autoencoder, significantly outperforming models trained on unaugmented data. This result demonstrates the effectiveness of combining deep learning with texture analysis and advanced augmentation strategies in handling complex veterinary diagnostic tasks. The findings highlight the potential of automated systems to improve early detection and accurate diagnosis of canine diseases, offering a reliable, scalable solution for real-world clinical applications.

data augmentation
local binary pattern (lbp)
autoencoder
yolov11
canine disease diagnosis
dermatological and ophthalmological imaging
cutmix
hybrid dataset
deep learning in veterinary medicine

2024 - International journal
(3)

A Fusion of RGB Features and Local Descriptors for Object Detection in Road Scene
Nguyen Dinh Vinh
Nguyen Dinh VinhCorresponding
Institute of Electrical and Electronics Engineers
Published: May 22, 2024

Many texture descriptors have been introduced in recent years to improve texture analysis and classification outcomes, which are important in many computer vision tasks including object recognition and detection, human detector, and especially in face recognition. Local pattern is a texture descriptor that can successfully extract distinctive texture features that possesses noise and illumination variance robustness. This paper focuses on making use of local pattern features in boosting object detection models in a multi-modal fusion paradigm to acquire reliable feature maps in forward propagation throughout the network regardless of variations in photo taking conditions. We propose an adaptive fusion architecture for RGB and Local Ternary Pattern information. This architecture leverage local pattern to enrich information of original feature maps and adapt to many object detection models. Our local pattern fusion network concentrates on backbone and neck modules with an simple and efficient operation. The notable accuracy advancement is 8.03% observed in Cascade R-CNN in KITTI Dataset. In difficult conditions, our fusion models significantly lift the original performance from 4.7% to 66.3% mAP score.

detectors
feature extraction
object detection
training
noise measurement
mathematical models
laser radar
multimodal sensors
An Experimental Comparison Between Tree-Based and Machine Learning Approaches for Predicting an Oil Temperature in an Electricity Transformer Dataset
Nguyen Dinh Vinh, Anh Kim Thi Do
Nguyen Dinh VinhCorresponding
Anh Kim Thi Do
Springer Nature Link
Published: August 31, 2024

Time series data forecasting remains a challenging problem within research communities. A plethora of machine learning and deep learning architectures have been devised to tackle the wide array of time series datasets across different domains. Recent advancements in machine learning algorithms have demonstrated impressive performance. However, there remains a limitation regarding the time-consuming nature of both training and testing phases. Hence, this study endeavors to address whether any methods can achieve comparable performance to recent machine learning algorithms without requiring extensive training time. Through our research, we have identified XGBoost as the optimal solution for addressing this issue. Through comprehensive experiments across three study cases encompassing multivariate to multivariate prediction, multivariate to univariate prediction, and univariate to univariate prediction scenarios, we discovered that XGBoost outperformed other state-of-the-art machine learning techniques in the multivariate-to-multivariate prediction case under ETD dataset.

deep learning
machine learning
performance comparison
time series forecasting
training time
testing time
xgboost
multivariate prediction
univariate prediction
etd dataset
state-of-the-art techniques
prediction scenarios
model efficiency
algorithm optimization
Enhancing semantic scene segmentation for indoor autonomous systems using advanced attention-supported improved UNet," in Signal, Image and Video Processing
V Hoang N. Tran, Nghi V. Nguyen, Nhi Q. P. Le, Nam N. N. Nguyen, Thu A. N. Le
V Hoang N. TranCorresponding
Nghi V. Nguyen
Nhi Q. P. Le
Nam N. N. Nguyen
Thu A. N. Le
Springer Nature Link
Published: December 11, 2024

This paper introduces EFFB7-UNet, an advanced semantic segmentation framework tailored for Indoor Autonomous Vision Systems (IAVSs) utilizing the U-Net architecture. The framework employs EfficientNetB4 as its encoder, significantly enhancing feature extraction. It integrates a spatial and channel Squeeze-and-Excitation (scSE) attention block, emphasizing critical areas and features to refine segmentation outcomes. Comprehensive evaluations using the NYUv2 Dataset and various augmented datasets were conducted. This study systematically compares EFFB7-UNet’s performance with multiple U-Net encoders, including ResNet50, ResNet101, MobileNet V2, VGG16, VGG19, and EfficientNets B0-B6. The findings reveal that EFFB7-UNet not only surpasses these configurations in terms of accuracy but also highlights the effectiveness of the scSE attention block in achieving superior segmentation results. Without the utilization of depth information, EFFB7-UNet achieves a 12% improvement in mean Intersection over Union (mIOU). This significant enhancement demonstrates EFFB7-UNet’s adaptability across various domains, implying substantial progress in enhancing the effectiveness and reliability of Intelligent Autonomous Vision Systems (IAVS) technologies.

segmentation
unet
semantic segmentation
attention mechanisms
convolutional neural networks (cnn)
efficientnet

2023 - International journal
(4)

A Robust Triangular Sigmoid Pattern-Based Obstacle Detection Algorithm in Resource-Limited Devices
Nguyen Dinh Vinh, T. D. Trinh, V Hoang N. Tran
Nguyen Dinh VinhCorresponding
T. D. Trinh
V Hoang N. Tran
Transactions on Intelligent Transportation Systems
Published: March 16, 2023

Object detection and classification are key processes in advanced driver-assistance systems. The existing object detection and classification methods are effective in normal daylight conditions. However, the performance of these methods deteriorates in adverse driving conditions, such as those involving low light, illumination changes, and nighttime conditions. To overcome these limitations, several feature-based algorithms have been developed that introduce local features, such as local binary pattern, local tetra pattern, and local density encoding, for adverse driving conditions. However, these local patterns cannot effectively address the noise in real driving conditions because the relationship between the neighboring pixels cannot be comprehensively encoded. To solve these problems, this study developed a robust feature-based method by introducing a triangular-pattern-based sigmoid function to effectively encode and establish the robust feature of neighboring pixels in the local region. The performance of the proposed pattern is evaluated by integrating it into state-of-the-art object detection algorithms. The proposed method significantly increases the vehicle detection ratio of YOLOv5s by 11.7% for an intersection over a union of 0.5 in difficult driving conditions for the CCD dataset. Moreover, the detection ratios of the proposed method are comparable to those of other state-of-the-art object detection methods such as Retina, Faster RCNN, and Deformable DETR over various datasets such as KITTI, COCO, HCI, and CCD. Additionally, the proposed algorithm is implemented on a Raspberry Pi-based autonomous car system to evaluate its performance during real driving conditions. Our proposed method supports robust input feature extraction and can thus be used to enhance the performance of the existing obstacle detection and classification systems.

feature extraction
object detection
roads
classification algorithms
lighting
charge coupled devices
performance evaluation
Robust Plant Leaves Diseases Classification Using EfficientNet and Residual Block
Nguyen Dinh Vinh, Q. N. Le, Narayan C. Debnath
Nguyen Dinh VinhCorresponding
Q. N. Le
Narayan C. Debnath
Springer Nature Link
Published: March 2, 2023

Vietnamese agriculture aims to develop sustainable intensification with the help of modern technology. Therefore, it is necessary to provide an intelligent system-based artificial intelligence to help farmers to analyze and diagnose the plant leaves disease to increase the plant health and productivity. Various systems have been proposed to detect and classify plant leaf diseases, such as convolutional neural networks (CNN), support vector machines (SVM), decision trees, KNN, and random forests. However, the performance of these existing systems is still not satisfactory for real applications. Therefore, we proposed a novel approach to accurately detect and classify plant leaf disease by using the EfficientNet and Residual block. The proposed system obtained better results than existing methods, such as CNN, SVM, KNN, random forest, and decision tree. The proposed system was able to detect and classify the blueberry healthy with a detection rate of 97.24%, the blueberry disease with the detection rate of 94.24%, the apple black rot with a detection rate of 95.14%, the tomato bacterial spot with a detection rate of 98.24%, and the potato late blight with the detection rate of 98.87%.

efficientnet
plant disease detection
image classification
artificial intelligence
residual block
vietnamese agriculture
smart farming.
Robust Vehicle Detection by Using Deep Learning Feature and Support Vector Machine
Nguyen Dinh Vinh, T. H. Tran, Dang Thai Doan, Narayan C. Debnath
Nguyen Dinh VinhCorresponding
T. H. Tran
Dang Thai Doan
Narayan C. Debnath
Springer Nature Link
Published: March 1, 2023

Vehicle detection and classification is an essential task in building an autonomous driving car. However, the accuracy of traditional vehicle detection systems is still not satisfied in real world due to unstable features from the original input images. This study introduces an efficient method to enhance the input feature more stable by considering the benefit of the machine learning approach to discover and extract robust features. The feature generated by the proposed deep learning model is then input to SVM to accurately identify objects. To evaluate the performance of the compared methods, we conducted comprehensive experiments on various datasets, such as KITTI, CCD, and HCI. For the KITTI dataset, the recall of our system, the LS-Support vector machine, the Linear support vector machine, the SVM-HOG, and the support vector machine are 71.4%, 64.3%, 53.8%, 67.8.0%, and 57.1%, respectively.

deep learning
feature extraction
performance evaluation
vehicle detection
machine learning
vehicle classification
autonomous driving
input feature enhancement
robust features
svm
object identification
kitti dataset
ccd dataset
hci dataset
ls-svm
linear svm
svm-hog
recall
real-world accuracy
Robust Vietnam’s Motorcycle License Plate Detection and Recognition Using Deep Learning Model
D. H. Le, D. Mazumder, L. D. Quach, S. Banerjee, Nguyen Dinh Vinh
D. H. LeCorresponding
D. Mazumder
L. D. Quach
S. Banerjee
Nguyen Dinh Vinh
Springer Nature Link
Published: November 17, 2023

Nowadays, driving fast is an emergency state in many countries in the world and Vietnam in particular, and the consequence is really cruel when an accident occurs. Therefore, it is very important to control the speed of vehicles. Motorcycle license plate recognition (MLPR) is one of many methods. If any motorcycle drives above the speed limit on the road that has a speed camera, a ticket will be sent to your living address. This research proposes a method that can automatically detect license plates (LP) and extract their data. The highest mAP achieved after three hundred epochs is 93% with Yolov8. There are three stages to extracting the LP’s data, the first is detecting the motorcycle, the second is detecting the LP in the motorcycle’s bounding box, and the third is LP recognition using Yolov8 also.

object detection
yolov8
accident prevention
speed control
motorcycle license plate recognition
mlpr
license plate detection
license plate recognition
vehicle monitoring
traffic enforcement
speed camera
map
automatic detection
lp data extraction
road safety
vietnam traffic

2023 - International conferences
(1)

Comprehensive Visual Features and Pseudo Labeling for Robust Natural Language-based Vehicle Retrieval
Bach Hoang Ngo, Dat Thanh Nguyen, Nhat-Tuong Do-Tran, Phuc Pham Huy Thien, Minh-Hung An, Tuan-Ngoc Nguyen, Loi Nguyen Hoang, Nguyen Dinh Vinh
Bach Hoang NgoCorresponding
Dat Thanh Nguyen
Nhat-Tuong Do-Tran
Phuc Pham Huy Thien
Minh-Hung An
Tuan-Ngoc Nguyen
Loi Nguyen Hoang
Nguyen Dinh Vinh
IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 2023 – AICity Challenge Track 2
Published: June 19, 2023

Vehicle retrieval has become crucial for public transportation and intelligent transportation systems due to the exponential development of large-scale transportation videos. Vehicle re-identification and vehicle tracking are the main components of most vision-based vehicle recovery systems. Unfortunately, the limited amount of information provided by traffic video feeds limits the vision-based vehicle retrieval algorithm’s efficacy. Therefore, this article proposes a contrastive cross-modal vehicle retrieval approach to maximize the complementarity of natural language and visual representations. An efficient method to fuse multiple input image features is also proposed to extract comprehensive information from various vehicles along with pseudo labeling and efficient post-processing techniques to enhance retrieval accuracy. The proposed method achieved the 3rd ranking of Mean Reciprocal Rank (MRR) score of 0.4795 on the test set for the Challenge Track 2: Tracked-Vehicle Retrieval by Natural Language Descriptions 2023. Source code for the proposed approaches is openly accessible at https://github.com/anminhhung/AICity-2023-Track2.

intelligent transportation systems
feature fusion
vehicle retrieval
large-scale transportation videos
vehicle re-identification
vehicle tracking
traffic video feeds
cross-modal retrieval
natural language
visual representation
pseudo labeling
post-processing
retrieval accuracy
mean reciprocal rank
mrr score
tracked-vehicle retrieval
aicity 2023
contrastive learning
computer vision
public transportation

2022 - International journal
(8)

A Deep Learning Framework for Robust and Real-Time Taillight Detection Under Various Road Conditions
H. -J. Jeon, Nguyen Dinh Vinh, T. T. Duong, J. W. Jeon
H. -J. JeonCorresponding
Nguyen Dinh Vinh
T. T. Duong
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: June 6, 2022

In this paper, we present a deep learning model for high-accuracy, high-speed detection of vehicle taillights in traffic. The model consists of three major modules: the lane detector, the car detector, and the taillight detector. Unlike most previously proposed algorithms where hand-coded schemes are used, we have adopted a data-driven approach. This data-driven scheme was implemented in both the car and taillight detection modules. First, we used an intricately designed lane detection module, then we adopted the Recurrent Rolling Convolution (RRC) architecture and tracking mechanism for detecting car boundaries. Subsequently, we used the same RRC architecture to extract the taillight regions of the detected cars. The lane detection and car detection modules improve both the speed and detection rate of the final taillight detection. The robustness of our model was verified using datasets from Sungkyunkwan University (SKKU) as well as the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI). Our model works well even in hostile conditions. It achieves detection rates as high as 99% in testing with the SKKU dataset. When using the KITTI 2D Object dataset, the model achieves a taillight detection rate of 86%. The model achieves 100% taillight detection rate on a certain, small subset of the KITTI Tracking dataset.

automobiles
pipelines
real-time systems
detectors
deep learning
support vector machines
lane detection
Face Mask Detection by Using Deep Learning with the Generative Adversarial Network-Based Feature
Nguyen Dinh Vinh, Narayan C. Debnath
Nguyen Dinh VinhCorresponding
Narayan C. Debnath
Springer Nature Link
Published: November 18, 2022

COVID-19 is one of the diseases that causes a lot of trouble in social communication. To prevent the spread of the Covid-19, people must wear a mask during conversation, shopping, or studying. Therefore, it is necessary to develop an application that helps to verify whether people wear a face mask or not. Many approaches have been proposed to detect a facemask based on using machine learning-based methods, such as support vector machines, and convolutional neural networks. However, the performance of existing systems still has limitations under difficult deployment environments where the camera's quality is not good enough for detection. Therefore, in this research, we study the benefits of generative adversarial networks to produce a stable feature for robust deep learning model-based facemask detection. First, the generative adversarial network is used to learn and discover stable features from the input dataset. Second, the Yolo is then employed to learn the stable feature to effectively detect and recognize facemasks under various testing conditions. With comprehensive experimental results, we found that our proposed method achieved a detection rate of 84.45% by using the MaskedFace-Net dataset.

deep learning
machine learning
covid-19
face mask detection
generative adversarial network (gan)
yolo
stable feature extraction
convolutional neural network (cnn)
support vector machine (svm)
maskedface-net dataset.
Face Recognition Based on Deep Learning and HSV Color Space
Nguyen Dinh Vinh, Q. N. Le, Narayan C. Debnath
Nguyen Dinh VinhCorresponding
Q. N. Le
Narayan C. Debnath
Springer Nature Link
Published: November 18, 2022

Face recognition has a wide range of applications from education to the industries. Existing face recognition systems were designed by using the RGB input only. Therefore, this research proposes a new system to increase the performance of the face recognition system by fusing the benefit of RGB, HSV, and raw RGB features. First, the HSV image was computed from the raw RGB input. Second, the LBP image was calculated from the draw RGB input. Third, the LBP, HSV, and RGB features are combined to produce a new feature. Fourth the proposed feature is then input to the deep learning-based FaceNet model to find the location of the human face. Finally, the support vector machine is used to identify the human name. Experimental results prove that our proposed system achieves high accurate detection rate. The proposed method improves the accuracy of existing face recognition systems by 2.15% in terms of the detection rate under our dataset and the Kaggle dataset. The proposed method obtained the detection rate of 96.25% on the custom dataset.

face recognition
deep learning
feature fusion
detection rate
support vector machine
rgb
hsv
lbp
facenet
human identification.
Plant Leaf Diseases Detection and Identification Using Deep Learning Model
D. H. Chau, T. H. Tran, Nguyen Dinh Vinh
D. H. ChauCorresponding
T. H. Tran
Nguyen Dinh Vinh
Springer Nature Link
Published: April 17, 2022

In agriculture, leaf diseases often appear in changing weather conditions. Changing weather conditions can be very rainy or very hot or very humid. These factors make plants susceptible to bacterial, fungal, and viral infections. This research investigates three common diseases on apple trees, such as black rot, fish scale disease, snow apple rust. We use the deep learning method-based Yolo-v5 model along with the proposed stable information based on auto-encoder to train a Plant-Village dataset with 5740 images; The proposed system uses Google Colab for training phases. The data set is divided into three parts: training, validation, and test. We used 70% dataset (4018 images) for training, used 20% dataset (1148) for performing the validation step, and 10% dataset (574 images) for a testing phase. After training, we obtained the result of 81.28% in terms of the detection rate and 91.93% in terms of the classification rate by using the Plant-Village dataset .

deep learning
agriculture
auto-encoder
image classification
leaf disease detection
apple tree diseases
yolov5
plant-village dataset
google colab
weather conditions
Real-Time Multi-Digit Recognition System Using Deep Learning on an Embedded System
H. -J. Jeon, J. W. Jeon, Nguyen Dinh Vinh
H. -J. JeonCorresponding
J. W. Jeon
Nguyen Dinh Vinh
Association for Computing Machinery
Published: July 21, 2022

In the machine learning technology, making computers read letters, characters and digits have been a hot issue in many areas of research. Among them, the recognition of handwritten numbers has still a long way to go, unlike the recognition of printed digits or the recognition of handwritten English sentences. In this paper we will introduce to you a multiple handwritten digit recognition system using deep learning.

deep learning
machine learning
pattern recognition
handwritten digit recognition
optical character recognition (ocr)
multiple digit recognition
handwritten character analysis
Robust Face Mask Detection Using Local Binary Pattern and Deep Learning
L. D. Quach, P. H. Tran, H. Q. D. Nguyen, Nguyen Dinh Vinh
L. D. QuachCorresponding
P. H. Tran
H. Q. D. Nguyen
Nguyen Dinh Vinh
Springer Nature Link
Published: April 17, 2022

The ongoing COVID-19 has caused a great amount of serious troubles for people around the world. Even though vaccination has been proven to be safe and highly effective against the COVID-19, it is far away to prevent thoroughly the spread of the disease and truly halt the pandemic. Therefore, we need to apply additional methods aside from vaccine injection, such as keeping the distance between people and always using the face masks during the ordinary conversations, in efforts to further reduce the COVID-19 contagion rate. To implement such methods, this research aims to investigate an efficient approach to detect and warn people that they should wear mask whenever they go to public places. Our proposed system studies the benefits of Local Binary Pattern (LBP) and deep learning model to provide accurate face mask detection and classification system. After comprehensive testing, we found that our system provided the detection rate up to 90% with the Kaggle, Face-Mask-Net, and our own datasets.

deep learning
covid-19
face mask detection
local binary pattern (lbp)
public health
mask classification
contagion prevention
kaggle dataset
face-mask-net dataset
image recognition.
Robust Median-Ternary Pattern for Traffic Light Detection
T. D. Trinh, D. D. Nguyen, T. P. Nguyen, D. H. Le, N. Van Nguyen, Narayan C. Debnath, Nguyen Dinh Vinh
T. D. TrinhCorresponding
D. D. Nguyen
T. P. Nguyen
D. H. Le
N. Van Nguyen
Narayan C. Debnath
Nguyen Dinh Vinh
Springer Nature Link
Published: May 19, 2022

Traffic light detection and classification still be challenging problems in driving assistance systems due to various unknown factors appearing along the driving roads, especially under night conditions. This research, therefore, introduces an efficient local feature that helps to make original feature input to be more robust under various driving conditions by investigating the benefits of ternary-encoding and global information. The proposed feature input is then transferred to the deep learning-based YOLO to detect and classify traffic light status. Experimental results show that the proposed method obtained stable performance with a detection rate of 73.18% and a classification rate of 92.22% under night conditions using the CCD dataset.

deep learning
yolo
traffic light detection
traffic light classification
driving assistance systems
night conditions
ternary encoding
global information
local features
ccd dataset.
Sigmoid Local Pattern for Robust Car and Pedestrian Detection
L. D. Quach, P. H. Nguyen, H. Q. D. Nguyen, Nguyen Dinh Vinh
L. D. QuachCorresponding
P. H. Nguyen
H. Q. D. Nguyen
Nguyen Dinh Vinh
Springer Nature Link
Published: May 19, 2022

In this paper, we propose a solution that can solve the problems of detecting cars and pedestrians under extremely difficult conditions, providing higher object detection and recognition rate. By analyzing the existing encoding-base feature method, we found that they do not consider the relationship between each channel when encoding the local region. The encoder was used to extract the information in the local region by considering the feature for each channel, separately. The fusion feature is combined by only considering the linear relationship between RGB input features and local features. This combination is not good enough because it is difficult to estimate the linear parameters during the running time. Therefore, it is necessary to introduce an efficient feature-based method that can help to resolve the current limitations of the existing methods. After analyzing the performance of the existing local pattern, we found that we can encode the local region by tolerating the relationship between the center features and neighbor features using an idea of a sigmoid activation function in the deep learning model. The proposed sigmoid-based method achieved the accuracy of 78% under extremely difficult conditions while the existing yolov5-based method achieved 65% under extremely difficult conditions.

deep learning
object detection
car detection
pedestrian detection
sigmoid activation
feature encoding
local pattern
rgb feature fusion
difficult conditions
yolov5 comparison

2021 - International journal
(6)

Adaptive ternary-derivative pattern for disparity enhancement
Nguyen Dinh Vinh, M. T. T. Nguyen, D. D. Nguyen, J. W. Jeon
Nguyen Dinh VinhCorresponding
M. T. T. Nguyen
D. D. Nguyen
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: February 13, 2021

10.1109/ICIP.2012.6467523

feature extraction
encoding
noise
brightness
belief propagation
measurement
standards
Feature Engineering and Deep Learning for Stereo Matching Under Adverse Driving Conditions
V. Q. Dinh, P. H. Nguyen, Vinh D. Nguyen
V. Q. DinhCorresponding
P. H. Nguyen
Vinh D. Nguyen
IEEE Transactions on Intelligent Transportation Systems
Published: April 28, 2021

Stereo matching is a challenging research topic in driving assistance systems. Existing stereo matching methods work well under normal day-light conditions. However, they fail to operate under adverse driving conditions, such as at night and during snowfall. This paper proposes a robust stereo matching framework using both deep-learning-based features and feature engineering. The proposed method investigates the benefits of features based on feature engineering and deep learning for solving stereo matching problems. Robust feature engineering is proposed for handling specific driving under adverse weather conditions, and a robust feature based on deep learning is considered for handling unspecific driving under extreme weather conditions. The proposed study has shown significantly improved accuracy by 8.31% for the state-of-the-art census based on semi-global matching under the reflection regions using the KITTI Stereo 2012 benchmark. Moreover, the experimental results demonstrate that the proposed system obtains more stable results than existing stereo methods based on deep learning on various stereo datasets, such as the Middlebury, EISAT, HCI, and CCD datasets.

real-time systems
deep learning
feature extraction
lighting
three-dimensional displays
intelligent transportation systems
correlation
Local Binary Pattern and Census, Which One is Better in Stereo Matching
Nguyen Dinh Vinh, P. H. Nguyen, Narayan C. Debnath
Nguyen Dinh VinhCorresponding
P. H. Nguyen
Narayan C. Debnath
Institute of Electrical and Electronics Engineers
Published: February 2, 2021

Local binary patterns and Census share similar ideas of encoding the local region by establishing the relationship between neighbor pixels to obtain robust feature transformation. Recently, LBP and its variants have been successfully applied in various applications, such as texture classification, face recognition, object detection, and segmentation, while Census has only been used to investigate stereo correspondence problem. Therefore, this paper investigates the LBP and Census using a non-local-based stereo matching method in order to analyze and discuss the main differences between LBP and Census. Moreover, as many as one hundred variants of LBP have been published to solve various problems, while only a few modifications of the Census exist for stereo matching. Comprehensive experiments with the indoor, Middlebury dataset stated that some novel LBPs that perform well in texture classification and face recognition also work well in a stereo matching application. In most cases, LBP and its variants compare favorably to Census in terms of the accuracy of the stereo method. These results proved that LBP and its variants are suitable for using in solving the stereo correspondence problem or improving the performance of existing stereo methods.

face recognition
object detection
encoding
pattern matching
data mining
hamming distance
measurement techniques
Local Tetra Pattern and Its Benefits to Improve the Performance of Car and Pedestrian Detection Under Hostile Conditions
A. L. Dang, T. Q. Nguyen, T. T. Cao, V. Q. Dinh, Nguyen Dinh Vinh
A. L. DangCorresponding
T. Q. Nguyen
T. T. Cao
V. Q. Dinh
Nguyen Dinh Vinh
Institute of Electrical and Electronics Engineers
Published: December 28, 2021

Traffic detection is a topic of great interest in recent years due to a high demand for better traffic detection systems. Existing traffic detection algorithms work well under ideal driving conditions, however their performance decreases under difficult conditions such as insufficient lighting and illumination. Recently, local patterns have been successfully applied in order to handle complex texture conditions, such as stereo matching, and texture classification. We propose a method that applies Local Tetra Pattern for data preprocessing, so as to improve the performance of deep learning models under said conditions. Our approach achieved better performance than the original raw-models while the changes in inference time are maintained within a negligible interval. By fusing local patterns and raw images, the model gains an acquisition of discriminative information in regions that are highly similar. In challenging conditions, these kinds of information are essential for the model to recover its consciousness of concerned objects which cause many re-cognitional obstructions. Experimental results show a percentage as high as 35.847%, an increase of 12.575% in comparison with the original result on the SKKU data set.

automobiles
deep learning
lighting
automation
data preprocessing
control systems
data models
Robust and Real-Time Deep Learning System for Checking Student Attendance
Nguyen Dinh Vinh, K. X. H. Nguyen Tran, V. C. Nguyen
Nguyen Dinh VinhCorresponding
K. X. H. Nguyen Tran
V. C. Nguyen
Journal of Advances in Information Technology.
Published: December 4, 2021

Abstract—A face detection and identification algorithm is an interesting research topic. The performance of existing face detection and identification systems works well under normal lighting conditions, while their performance is not stable under difficult conditions due to noise and illumination changes. Therefore, this research aims to develop a robust and real-time deep learning system for student face detection and identification to overcome these current limitations. The proposed method investigates both benefits of state-of-the-art deep learning models and local patterns to create a robust frame-work for detecting and checking student attendance. Comprehensive experimental results show that the proposed method obtained stable results under various normal and difficult indoor conditions. The proposed method obtains the detection rate of 93.55% and 89.25% under normal and difficult indoor conditions, respectively. The proposed method obtains the identification rate of 87.79% and 85.19% under normal and difficult indoor conditions, respectively.

face recognition
deep learning
face detection
attendance system
local binary pattern
multiple features
Robust Crosswalk Detection Using Deep Learning Approach
T. D. Trinh, D. D. Nguyen, T. P. Nguyen, T. N. D. Le, N. Van Nguyen, Narayan C. Debnath, Nguyen Dinh Vinh
T. D. TrinhCorresponding
D. D. Nguyen
T. P. Nguyen
T. N. D. Le
N. Van Nguyen
Narayan C. Debnath
Nguyen Dinh Vinh
Springer Nature Link
Published: November 9, 2021

The driving assistance system is going to be true shortly. Therefore, developing a robust sub-component system, such as crosswalk detection, is necessary to support the drivers and prevent accidents. Existing crosswalk detection has been studied by using engineering feature-based methods and machine learning-based methods. However, the performance of existing systems based on engineering and machine learning still has several limitations. Therefore, this paper introduces a new approach to improving existing crosswalk detection performance by investigating the benefits of the data augmentation, unsupervised autoencoder, and state-of-the-art YOLO deep learning method. First, the input image is converted to HSV color space, LAB color space, and LUV color space. Second, the HSV, LAB, and LUV color results and raw RBC are then concatenated at the next phase. Thus, the size of the input feature is increased because we join the multiple input features. Therefore, we proposed an approach to use an auto-encoder to reduce the input dimension of the proposed system. The result yields that our proposed crosswalk detection got the detection rate and false alarm rate of 85.45% and 2.5 % under the SKKU dataset, respectively. And, we also got the detection rate and false alarm rate of 88.15% and 1.45 % under our custom fly cam dataset.

feature fusion
data augmentation
yolo
driving assistance systems
crosswalk detection
autoencoder
hsv color space
lab color space
luv color space
skku dataset
fly cam dataset.

2020 - International journal
(2)

Robust and Real-Time Obstacle Region Detection Based on Depth Feature for Vehicle Detection
P. M. Nguyen, V. C. Nguyen, Nguyen Dinh Vinh
P. M. NguyenCorresponding
V. C. Nguyen
Nguyen Dinh Vinh
Springer Nature Link
Published: March 24, 2020

Obstacle region detection has been one of the most important tasks for detecting and classifying preceding vehicles moving along the road. Traditional U-V disparity-based obstacle detection methods consider only single disparity pixels to extract obstacle regions. They might fail to produce accurate obstacle regions when noise is encountered in the original disparity map. Recently, local patterns have been successfully applied to handle noise in many systems, such as texture classification and vehicle detection. Therefore, this paper introduces a new approach to compute U and V disparity using maximum local density encoding and ternary pattern features. Our method produced better performance than the existing U-V disparity methods in terms of detection rate and accuracy. In addition, we also evaluated the performance of the proposed method with the Faster RCNN-based object detection. Experimental results show that our method improved the detection rate of Faster RCNN by 1.65% on the KITTI dataset, and by 2.28% on the CCD dataset. In addition, the proposed method also improved the running time of Faster RCNN by 47% on the KITTI dataset.

vehicle detection
kitti dataset
ccd dataset
obstacle detection
u-v disparity
local density encoding
ternary pattern
faster rcnn
disparity map
noise robustness
Robust Traffic Light Detection and Classification Under Day and Night Conditions
P. M. Nguyen, V. C. Nguyen, Nguyen Dinh Vinh
P. M. NguyenCorresponding
V. C. Nguyen
Nguyen Dinh Vinh
International Conference on Intelligent Information Technology
Published: December 1, 2020

Recently, traffic light detection and classification systems have been studied and developed to build an autonomous car by many research institutes, universities, and companies. However, the results of existing traffic light detection systems are still not stable under day and night conditions. It is difficult to detect the location of traffic light due to their small size. Moreover, traffic lights’ shapes are also similar to advertisement lights in a city road. Therefore, this paper proposed a new approach to improve the performance of existing traffic light detection systems by using the benefits of hand-crafted features and deep learning techniques. Experimental results show that the proposed system obtained the detection rate of 80% under night conditions, while the color-based density method only got the detection rate of 50.43% under night conditions.

real-time systems
deep learning
feature extraction
training
charge coupled devices
image color analysis
colored noise

2018 - International journal
(2)

Real-Time Vehicle Detection Using an Effective Region Proposal-Based Depth and 3-Channel Pattern
Nguyen Dinh Vinh, J. J. Byun, D. T. T., J. W. Jeon
Nguyen Dinh VinhCorresponding
J. J. Byun
D. T. T.
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: November 14, 2018

Traditional deep learning-based vehicle detection methods are often designed using a pyramid of filters with multiple scales and sizes; therefore, the processing time is slow due to the large number of scales used and because the classifier runs at all scales. Recently, a deep learning-based region proposal network was introduced to detect vehicles that only employ the network one time regardless of the size of the input image. In object detection, deep learning-based region proposal networks have achieved state-of-the-art performance in terms of accuracy. These systems achieve a very high accuracy under normal driving conditions; however, their performance decreases under difficult driving conditions such as in snow, rain, or fog. In addition, the current state-of-the-art system-based region proposal networks still fail to satisfy the real-time requirements of the driving assistant systems. More recently, the identification of local patterns has been shown to improve the performance of the traditional deep-learning systems; hence, this paper investigates local patterns in region proposal networks to improve their accuracy. Depth information is also investigated to improve the processing time of current region proposal networks. Our experimental results show that the proposed system obtains better performance than the state-of-the-art object region detection systems in terms of both accuracy and processing time.

real-time systems
object detection
proposals
vehicle detection
neural networks
cameras
Robust Pedestrian Detection via a Recursive Convolution Neural Network
D. T. Tran, Nguyen Dinh Vinh, J. W. Jeon
D. T. TranCorresponding
Nguyen Dinh Vinh
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: August 23, 2018

Pedestrian detection is fundamental challenge for computer vision which requires localizing objects within an image. Convolutional neural networks are widely used in object recognition. However, ordinal convolutional methods using sliding window as the input for networks require time to run an entire image and can only handle a fixed size window image. We propose to using a region proposal based in the V-disparity method to obtain prospect regions, instead of the original scanning methods to obtain the object regions. The region proposals from the V-disparity will be fed as the input for a convolutional neural network(CNN). We also extend the CNN for more effective task object detection. In our model, CNN are combined with recursive neural networks to learn features and classify color images. The convolutional neural network layer learns low-level features of the input image. By using CNN, the learned features can represent highly variable objects in the input image. The learned features after the convolutional layer are then given as inputs to a recursive neural network (RNN) to compose higher order features. The RNN is a multiple, fixed-tree recursive which can combine convolution and pooling into one efficient hierarchical operation.

feature extraction
proposals
computer vision
convolutional neural networks
task analysis
microsoft windows

2017 - International journal
(3)

High-Speed Car Detection Using ResNet-Based Recurrent Rolling Convolution
H. -J. Jeon, C. C. Pham, Nguyen Dinh Vinh, J. W. Jeon
H. -J. JeonCorresponding
C. C. Pham
Nguyen Dinh Vinh
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: January 19, 2017

Car detection is a crucial issue in self-driving cars. Numerous in-traffic car detection models have been proposed, each of which exhibits its own strengths and weaknesses; the high detection speeds of some models are not accompanied by high precision, while the precision of other models is shadowed by insufficient speeds. Our main goal in this paper is to introduce a model that utilizes the Recurrent Rolling Convolution (RRC). The model gives promising results on detection speed and precision, thereby mitigating the weaknesses of previously proposed models, which is exhibited in our extensive experiment.

automobiles
detectors
feature extraction
object detection
data models
convolution
computational modeling
Pedestrian detection based on deep learning model
H. -J. Jeon, Nguyen Dinh Vinh, J. W. Jeon
H. -J. JeonCorresponding
Nguyen Dinh Vinh
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: February 16, 2017

Pedestrian detection remains an important task in the theory research and practical application of objects detection. Traditional pedestrian detection algorithms require experts design features to describe the pedestrian characteristics and combine with the classifiers. In recent years, deep learning and especially Convolutional Neural Networks (CNN) have made great success on image and audio, which is the important component of deep learning. Artificial designed methods of feature extracting has an imperfect description of pedestrian in the complex background. In this paper, we propose a pedestrian detection method based on deep convolutional neural network with multi-layers. It can make full use of the advantages of deep convolutional neural network and extract features from the database of pedestrian detection. At the stage of region proposal, to solve the problem of too much redundant windows generated by traditional methods, we use the edge boxes algorithm instead of sliding window algorithm to extract windows. At last, we get a smaller number of windows with high-quality, which is of great importance for the subsequent classification task. At the end of the paper, we carried out multi-sets of comparison experiments in this system. Experiments show that the pedestrian detection system based on deep learning outperforms the traditional methods based on both handcrafted and learned features.

feature extraction
classification algorithms
proposals
neural networks
machine learning
image edge detection
convolution
Real-time pedestrian detection using a support vector machine and stixel information
M. T. T. Nguyen, Nguyen Dinh Vinh, J. W. Jeon
M. T. T. NguyenCorresponding
Nguyen Dinh Vinh
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: December 14, 2017

Pedestrian detection systems have become a crucial topic due to their contribution to numerous applications, such as surveillance, self-driving cars, tracking systems, and robotics. Therefore, we have developed an efficient pedestrian detection algorithm using a support vector machine (SVM) and a histogram of oriented gradients (HOG) feature. This paper introduces a system that is capable of reducing the set of candidate detection regions and classifying pedestrians and non-pedestrians in urban traffic by using the stixel world and HOG + SVM, respectively. The stixel world computation in this proposal is used to compute the region of interest of the input image. We also introduce our new human dataset, including more than 3100 human images, and a method used for classifying the dataset. By implementing on a CPU only, we reach high-quality detection at 56 fps in urban traffic scenarios. Our experimental results demonstrate the effectiveness of reducing the set of detection regions to improve the processing time of our system.

support vector machines
feature extraction
data mining
brightness
estimation
detection algorithms
histograms

2016 - International journal
(2)

Learning Framework for Robust Obstacle Detection, Recognition, and Tracking
Nguyen Dinh Vinh, H. V. Nguyen, D. T. T., J. W. Jeon
Nguyen Dinh VinhCorresponding
H. V. Nguyen
D. T. T.
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: October 19, 2016

This paper introduces a general framework for detection, recognition, and tracking preceding vehicles and pedestrians based on a deep learning approach. The proposed framework combines a novel deep learning approach with the use of multiple sources of local patterns and depth information to yield robust on-road vehicle and pedestrian detection, recognition, and tracking. The proposed system is first based on robust obstacle detection to identify obstacles appearing along the road that are likely to be vehicles and pedestrians, implemented as an efficient adaptive U-V disparity algorithm. Second, the results from the obstacle detection stage are input into a novel vehicle and pedestrian recognition system based on a deep learning model that processes multiple sources of depth information and local patterns. Finally, the results from the recognition stage are used to track detected vehicles or pedestrians in the next frame by means of a proposed tracking and validation model. The proposed framework has been thoroughly evaluated by inputting several vehicle and pedestrian data sets that were collected under various driving conditions. Experimental results show that this framework provides robust vehicle and pedestrian detection, recognition, and tracking with high accuracy, and also satisfies the real-time requirements of driver assistance systems.

training
roads
neural networks
cameras
vehicles
machine learning
robustness
Robust Stereo Data Cost With a Learning Strategy
Nguyen Dinh Vinh, J. W. Jeon
Nguyen Dinh VinhCorresponding
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: August 24, 2016

The performance of stereo matching algorithms strongly depends on the quality of the stereo data/matching cost. Most state-of-the-art data costs require expert knowledge for the design of a transformation function, such as census for handling gray-level changes monotonically, adaptive normalized cross correlation for handling Lambertian cases, guided filtering for preserving edge information, and local density encoding for handling illumination differences. However, it is difficult to design a complex transformation function to handle unknown factors that often occur in driving conditions such as snow, rain, and sun. Therefore, this paper has investigated the deep learning strategy to develop a novel stereo matching cost model without using much expert knowledge. Experimental results show that the proposed deep learning model obtains better results than the state-of-the-art stereo matching cost as judged by the standard KITTI benchmark, Middlebury, and HCI datasets.

feature extraction
training
machine learning
robustness
image color analysis
algorithm design and analysis
snow

2015 - International journal
(2)

Fuzzy Encoding Pattern for Stereo Matching Cost
V. Q. Dinh, Nguyen Dinh Vinh, J. W. Jeon
V. Q. DinhCorresponding
Nguyen Dinh Vinh
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: July 13, 2015

We propose a novel fuzzy encoding pattern that fuzzily encodes the relative orders between pixel pairs. An image window is divided into disjoint neighboring pixel sets for the window's center pixel, and the relative order is established not only between the center pixel and its neighbors but also between the pixel pairs in each neighboring pixel set. The relative orders are fuzzily encoded to extract more detailed information from a local structure. We successfully apply the pattern as a matching cost function for stereo correspondence under severe radiometric variations. We conduct experiments using the proposed matching cost function and compare it with functions employing the census transform, supporting local binary pattern, and adaptive normalized cross correlation, as well as a mutual information-based matching cost function, using different stereo data sets. Compared with the census transform, the proposed function reduces the error from 33.1% to 16.9% in the Middlebury data set and from 17.6% to 9.5% in the Kitti data set. The experimental results indicate that the proposed function is superior to the state-of-the-art functions under radiometric variations. In addition, the proposed function is faster than recently developed functions, such as the adaptive normalized cross correlation, a mutual information-based function, and support local binary pattern.

cost function
radiometry
transforms
encoding
distortion
pattern matching
data mining
Robust Matching Cost Function for Stereo Correspondence Using Matching by Tone Mapping and Adaptive Orthogonal Integral Image
V. Q. Dinh, Nguyen Dinh Vinh, J. W. Jeon
V. Q. DinhCorresponding
Nguyen Dinh Vinh
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: September 23, 2015

Real-world stereo images are inevitably affected by radiometric differences, including variations in exposure, vignetting, lighting, and noise. Stereo images with severe radiometric distortion can have large radiometric differences and include locally nonlinear changes. In this paper, we first introduce an adaptive orthogonal integral image, which is an improved version of an orthogonal integral image. After that, based on matching by tone mapping and the adaptive orthogonal integral image, we propose a robust and accurate matching cost function that can tolerate locally nonlinear intensity distortion. By using the adaptive orthogonal integral image, the proposed matching cost function can adaptively construct different support regions of arbitrary shapes and sizes for different pixels in the reference image, so it can operate robustly within object boundaries. Furthermore, we develop techniques to automatically estimate the values of the parameters of our proposed function. We conduct experiments using the proposed matching cost function and compare it with functions employing the census transform, supporting local binary pattern, and adaptive normalized cross correlation, as well as a mutual information-based matching cost function using different stereo data sets. By using the adaptive orthogonal integral image, the proposed matching cost function reduces the error from 21.51% to 15.73% in the Middlebury data set, and from 15.9% to 10.85% in the Kitti data set, as compared with using the orthogonal integral image. The experimental results indicate that the proposed matching cost function is superior to the state-of-the-art matching cost functions under radiometric variation.

lighting
robustness
cost function
radiometry
nonlinear distortion
transforms

2014 - International journal
(1)

Local Density Encoding for Robust Stereo Matching
Nguyen Dinh Vinh, D. D. Nguyen, S. J. Lee, J. W. Jeon
Nguyen Dinh VinhCorresponding
D. D. Nguyen
S. J. Lee
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: June 30, 2014

Stereo correspondence is challenging under realistic conditions due to uncontrolled factors that affect input images, including illumination inconsistencies and radiometric variations. Many local and global models have been suggested to address these problems; however, their performance is often degraded due to the assumption of color consistency between the left and right images. Therefore, we present a new local pattern, local density encoding, for stereo matching measurements to improve the performance of existing stereo methods. Our experimental results indicate that the proposed method is less sensitive to illumination changes and radiometric variations. Moreover, in the cases with normal and severe illumination changes, the proposed method is more robust than state-of-the-art data costs.

lighting
robustness
radiometry
encoding
noise
brightness
rain

2013 - International journal
(3)

A Fast Evolutionary Algorithm for Real-Time Vehicle Detection
Nguyen Dinh Vinh, T. T. Nguyen, D. D. Nguyen, S. J. Lee, J. W. Jeon
Nguyen Dinh VinhCorresponding
T. T. Nguyen
D. D. Nguyen
S. J. Lee
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: January 25, 2013

The evolutionary algorithm (EA) is an effective method for solving various problems because it can search through very large search spaces and can quickly come to nearly optimal solutions. However, existing EA-based methods for vehicle detection cannot achieve high performance because their fitness functions depend on sensitive information, such as edge or color information on the preceding vehicle. This paper focuses on improving the performance of existing evolutionary-based methods for vehicle detection by introducing an effective fitness function that can more accurately capture a vehicle's information by combining a disparity map, edge information, and the position and motion of the preceding vehicle. The proposed method can detect multiple vehicles by using a turn-back genetic algorithm (GA) and can prevent false detection by using motion detection. Our fitness function is designed in a typical manner along with the fitness parameters. These parameters are usually selected using heuristic methods, making the choice of optimal parameters difficult. Therefore, this paper proposes a new approach to estimating optimal fitness parameters using EA and the least squares method. Robustness testing showed that the proposed method provides detection rate (DR) results close to those obtained using a state-of-the-art system and outperforms other dominant vehicle-detection-based EAs.

feature extraction
vehicle detection
cameras
vehicles
biological cells
image edge detection
stereo vision
Efficient spatio-temporal local stereo matching using information permeability filtering
C. C. Pham, Nguyen Dinh Vinh, J. W. Jeon
C. C. PhamCorresponding
Nguyen Dinh Vinh
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: February 21, 2013

Spatiotemporal stereo matching has attracted much interest over the last few years. The problem is to maintain the temporally consistent disparity of video sequences and to avoid the flickering-artifacts occurring in consecutive disparity maps. In this paper, we propose a constant time spatiotemporal local stereo matching based on the information permeability method that was recently proposed and achieves good stereo results for static image pairs. The proposed three-pass aggregation method takes into account multiple preceding and following frames of the video instead of a single image pair as in the conventional method. Consequently, the temporal disparity consistency is enforced without requiring an explicit motion estimation as several spatiotemporal approaches have done. Experiments showed that our method improves the disparity consistency compared to the conventional information permeability method, and that it outperforms existing spatiotemporal stereo matching techniques.

spatiotemporal phenomena
permeability
smoothing methods
ip networks
filtering
integrated optics
optical imaging
Support Local Pattern and its Application to Disparity Improvement and Texture Classification
Nguyen Dinh Vinh, T. T. Nguyen, D. D. Nguyen, V. Q. Dinh, J. W. Jeon
Nguyen Dinh VinhCorresponding
T. T. Nguyen
D. D. Nguyen
V. Q. Dinh
J. W. Jeon
IEEE Transactions on Intelligent Transportation Systems
Published: March 27, 2013

The local binary pattern (LBP) and its variants have been widely investigated in many image processing and computer vision applications due to their robust ability to capture local image structures and their computational simplicity. The existing LBPs extract local structure information by establishing a relationship between the central pixel and its adjacent pixels. However, most LBPs miss the relationship among all of the pixels in the local region. Therefore, this paper proposes a novel model to establish this relationship by introducing a support LBP. The proposed model improves the performance of the existing LBP methods and results in lower sensitivity to illumination changes and radiometric variations. Moreover, the proposed model has been successfully investigated in two applications: disparity map generation and texture classification. For disparity map generation, the proposed model reduces the root mean square (RMS) error by 23.6% (in Baby1 dataset, Middlebury), and 16.58% (in Aloe dataset, Middlebury) as compared with the standard LBP under radiometric variation conditions. Moreover, the proposed model reduces the RMS by 28.11% as compared with the standard LBP under the Gaussian noise condition in the ESATS dataset. For texture classification applications, the proposed model improves the classification results from 96.26% to 98.13% on the Outext database, from 88.03% to 91.41% on the Xu database, and from 94.00% to 96.67% on the KTH-TIPS database as compared with the completed LBP.

binary sequences
pattern recognition
image processing

2012 - International journal
(3)

FPGA design and implementation of a real-time vehicle detection system
J. D. Jin, Nguyen Dinh Vinh, S. J. Lee, J. W. Jeon
J. D. JinCorresponding
Nguyen Dinh Vinh
S. J. Lee
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: December 31, 2012

In this paper, we propose a real-time vehicle detection system based on stereo vision system. Our system is composed of pipelined parallel processing techniques based on dedicated hardware architecture. The entire vehicle detection process, including histogram equalization, Canny edge detection, vehicle corner detection, vehicle region expansion, and vehicle validation, is made using a single field-programmable gate array (FPGA) without any external devices. Our system is about 3 times faster than a software program on a computer, and it shows high accuracy and low false rate. Thus, our proposed system is suitable for real-time processing.

real-time systems
vehicle detection
vehicles
image edge detection
histograms
hardware
field programmable gate arrays
Local stereo matching using an variable window, census transform and an edge-preserving filter
V. Q. Dinh, D. D. Nguyen, Nguyen Dinh Vinh, J. W. Jeon
V. Q. DinhCorresponding
D. D. Nguyen
Nguyen Dinh Vinh
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: December 31, 2012

In this paper, we propose an alternative blocking-matching approach to the correspondence problem in stereo matching. In blocking-matching algorithms, a local window is used to measure the similarity (or dissimilarity) between pixels of a stereo pair. Although some area-based stereo matching methods have been developed and work well in many kinds of regions such as textureless or object boundary regions, their performance can degrade when working in some types of radiometric conditions. Our proposed algorithm is an improved method that uses a non-parametric transform in the pre-processing step and an edge-preserving filter in the post-processing step. Input images are first pre-processed by the census transform, which makes the proposed method more robust when the image pair is captured in different light sources or camera exposure conditions. The window cost in our approach is computed from the transformed images using the Hamming distance, and the correspondence is finally chosen by a Winner-Takes-All strategy. The experimental results for the Middleburry images show that the proposed method outperforms test local stereo methods in radiometrically different images.

lighting
image color analysis
radiometry
transforms
computational efficiency
shape
measurement errors
Toward Real-Time Vehicle Detection Using Stereo Vision and an Evolutionary Algorithm
Nguyen Dinh Vinh, M. T. T. Nguyen, D. D. Nguyen, J. W. Jeon
Nguyen Dinh VinhCorresponding
M. T. T. Nguyen
D. D. Nguyen
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: July 16, 2012

A new approach for vehicle detection and distance estimation based on stereo vision and evolutionary algorithm (SEA) is described in this paper. First, we reuse our recent work on FPGA implementation of census-based correlations for stereo matching. Next, the SEA uses the gray scale left image and disparity information obtained from the FPGA system to detect the preceding vehicle and estimate its distance. This paper introduces an effective fitness function that allows our proposed method to have an improved performance and higher accuracy when compared with the existing evolutionary algorithm (EA) based methods. A new crossover type, tourna-ment crossover, is introduced to reduce the convergence time of our proposed. This paper also introduces a new approach for estimating the fitness function parameters. This estimation differs from the traditional EA because these parameters were generally created via experiments. Moreover, the processing time and accuracy of SEA can be improved by converting the global search to the local search with V disparity map. The robust experiments have proved that SEA successfully detects vehicles in front and sustains noise from different objects appearing along the road. The detection range is 10m-140m, the detection rate is 95% and the average processing-time is approximately 31 ms/frame on CPU. These results prove that SEA is suitable for a real-time system.

roads
vehicle detection
vehicles
biological cells
real time systems
genetic algorithms
convergence

2011 - International journal
(2)

Hardware architecture design for vehicle detection using a stereo camera
J. D. Jin, D. K. Kim, J. H. Seo, Nguyen Dinh Vinh, J. W. Jeon
J. D. JinCorresponding
D. K. Kim
J. H. Seo
Nguyen Dinh Vinh
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: December 19, 2011

Vehicle detection systems are increasingly important in providing vehicle road safety. However, many problems remain, such as a limit in the number of vehicles detected, false vehicle detection, and large processing time. Thus, we propose an efficient method for vehicle detection. This paper simulates vehicle detection using a rectified image and a disparity image based on dedicated hardware architecture. Every vehicle detection function, including vehicle corner detection, vehicle region expansion, and vehicle region validation, is implemented using parallel architecture. Our proposed hardware architecture is suitable for real time processing using the dedicated hardware architecture. In addition, we can detect many vehicles in the rectified image and reduce the false detection of vehicles.

vehicle detection
vehicles
image edge detection
stereo vision
real time systems
hardware
computer architecture
Real-time vehicle detection design and implementation on GPU
Nguyen Dinh Vinh, M. T. T. Nguyen, D. D. Nguyen, J. W. Jeon
Nguyen Dinh VinhCorresponding
M. T. T. Nguyen
D. D. Nguyen
J. W. Jeon
Institute of Electrical and Electronics Engineers
Published: November 19, 2011

Vehicle detection and distance estimation system has become important due to their assistance in reducing vehicle accidents. Therefore, an efficient vehicle detection and distance estimation algorithm using a knowledge-based method and image segmentation technique has been developed. The proposed algorithm can detect and estimate the distance of the preceding vehicle under various road conditions using a single CCD camera of 16 mm and 25 mm focal lenghts mounted on a vehicle. A GPU implementation of this proposed algorithm is introduced to enable our proposed system to support real-time processing. Experimental results under various road and weather conditions prove that our proposed system is suitable for a real-time system.

roads
vehicle detection
vehicles
image edge detection
real time systems
graphics processing unit
image segmentation

We use necessary cookies to run this site and optional analytics cookies to understand how it is used. Analytics only loads if you accept. Read our cookie policy