An Efficient and Cost-Effective Vehicle Detection and Tracking System for Collision Avoidance in Foggy Weather

An Efficient and Cost-Effective Vehicle Detection and Tracking (ECE-VDT) system is proposed and is essential for real-time collision avoidance and driver assistance in foggy weather conditions. The proposed ECE-VDT system encompasses vehicle detection and tracking algorithms. The optimized SimYOLO-V5s algorithm is proposed for vehicle detection, and the optimized deep-SORT algorithm is proposed for vehicle tracking in foggy weather. Several design changes are considered to optimize the baseline YOLO-V5s algorithm. The backbone network of baseline YOLO-V5s is optimized by integrating the SimSPPF module, a variant of the standard SPPF module. Furthermore, diverse distance metrics and localization/regression loss functions, such as CIOU, DIOU, EIOU, GIOU, and SIOU, are considered to enhance the baseline YOLO-V5s bounding box localization accuracy. These design changes significantly impacted the baseline YOLO-V5s vehicle detection performance. The baseline deep-SORT algorithm is optimized by utilizing the Sigmoid Linear Unit (SiLU) activation function in the pre-trained CNN network compared to the Rectified Linear Unit (ReLU) activation function. The self-annotated and customized state-of-the-art foggy weather vehicle detection datasets, Vehicle Detection in AdverseWeather Nature (DAWN), Foggy Driving (FD), and Foggy Cityscapes (FC), are considered for evaluating the efficient and cost-effective performance of vehicle detection in adverse weather conditions. Extensive and rigorous experiments on cloud and local GPU-enabled systems show that the optimized SimYOLO-V5s variants, more specifically SimYOLO-V5s_GIOU outperformed in mAP50 multi-class vehicle detection by 1.6 % on the DAWN dataset, 15.2 % on the FD dataset, and 3.4 % on the FC dataset compared to baseline YOLO-V5s. The optimized SimYOLO-V5s variants also outperformed state-of-the-art algorithms in accuracy and speed. The optimized deep-SORT_SiLU outperforms pre-processing, inference,NMSpost-processing, and update time, achieving higher FPS on foggy video sequences and the BDD100K dataset compared to the baseline deep-SORT and strong-SORT vehicle tracking algorithms. The proposed ECE-VDT system is an efficient and cost-effective solution to assist human-centric drivers of low-end vehicles and autonomous vehicles to avoid road collisions in foggy weather.