Advanced Vehicle Tracking System: Leveraging Gabor Feature Extraction and YOLOv5-Deep SORT for Real-Time MultiClass Traffic Monitoring.

Abstract

Listen

Translate article

The rapid growth of urban traffic necessitates efficient and accurate vehicle tracking systems to manage congestion, ensure safety, and support intelligent transportation systems. This study presents an Advanced Vehicle Tracking System that integrates Gabor feature extraction with the YOLOv5-Deep SORT framework for real-time, multi-class traffic monitoring of vehicles, specifically focusing on cars, buses, and trucks. By leveraging the strengths of Gabor filters in feature extraction and the robust detection capabilities of YOLOv5 coupled with Deep SORT for tracking, the system addresses key challenges of occlusion, classification errors, and real-time performance in dense traffic environments. The research builds on existing literature by enhancing feature extraction accuracy and integrating state-of-the-art detection and tracking models. A unique contribution of this work is the integration of Gabor filters to refine vehicle feature extraction, reducing misclassification and improving tracking stability. The methodology was validated using a custom dataset comprising annotated images from France, Belgium, Switzerland, Malaysia, and Nigeria, ensuring diversity in traffic scenarios. The dataset included 10,000 images with a balanced representation of the three vehicle classes. Experimental results indicate that the proposed system achieved an average precision (AP) of 94.2%, with recall and F1-scores exceeding 91% for all vehicle classes. The integration of Gabor features improved detection precision by 5% compared to baseline YOLOv5 models, particularly in cases of partial occlusion. Real-time performance metrics revealed an average processing time of 25 ms per frame, meeting the requirements for real-time monitoring applications. This research fills a critical gap in developing a multi-class vehicle tracking system capable of handling diverse traffic conditions and occlusions. While previous studies primarily focused on singleclass detection or struggled with occlusion challenges, the proposed system delivers superior accuracy and robustness. Future work will explore the integration of transformer-based models for further improvement in feature representation and the use of advanced optimization techniques to reduce computational overhead. Additionally, extending the system to include other vehicle classes, such as motorcycles and bicycles, will broaden its applicability. Collaboration with transportation agencies to deploy and test the system in real-world scenarios is recommended to validate its scalability and effectiveness.