International Journal of Sustainable Infrastructure for Cities and Societies

Traffic Surveillance Systems through Advanced Detection, Tracking, and Classification Technique

Aravind Sasidharan Pillai

Principal Data Architect, Data Engineering, Cox Automotive Inc., USA. ORCID: 0000-0001-7139-2804. Alumnus, Master in Data Science, University of Illinois Urbana-Champaign

https://orcid.org/0000-0001-7139-2804

Keywords: Advanced Data Augmentation, Traffic Management, Vehicle Detection, Yolov5 Model, Environmental Variability, Hyperparameter Sensitivity, Annotation Techniques

Abstract

Background: Traffic management and control are critically dependent on effective vehicle flow processing, including counting and tracking. Traditional methods often fall short in complex scenarios involving brightness changes and partial occlusions. This research addresses these challenges by implementing the YOLOv5 model, leveraging its robust architecture and advanced data augmentation techniques for improved vehicle detection and counting in diverse conditions.

Methodology: The YOLOv5 model is central to our approach, featuring a New CSP-Darknet53 backbone for feature extraction, SPPF and New CSP-PAN in the neck for feature integration, and a YOLOv3 inspired head for output generation. These components collectively enhance the model's sensitivity and accuracy in vehicle detection across varying scenarios. Data augmentation strategies such as Mosaic, Copy-Paste, and MixUp augmentations play a crucial role in preparing the model for real-world complexities. Furthermore, strategies like multiscale training and AutoAnchor optimization are employed to refine the detection and tracking process. The study also explores various annotation techniques and tools, including OpenCV and Numpy, to aid in the meticulous annotation process required for training and evaluation.

Experimentation: Our experiments utilize NVIDIA Tesla T4 GPUs, assessing the system's performance across several metrics, including precision, recall, F1 score, and more. Results indicate high precision (92%) and recall (88%), with an overall accuracy of 91%. The system demonstrates good data efficiency and robustness in varied conditions, though it shows sensitivity to hyperparameter settings. The research highlights the potential of YOLOv5 in improving traffic surveillance systems through enhanced detection, tracking, and classification capabilities.

Conclusion: The integration of the YOLOv5 model, coupled with advanced data augmentation and annotation strategies, offers significant improvements in vehicle detection and tracking. While challenges remain, particularly in handling occlusions and environmental variability, our findings suggest that with careful tuning and optimization, YOLOv5 can be a valuable tool in the advancement of traffic management systems.