A Deep Learning Based Real-Time Vehicle Detection and Collision Avoidance System for Low-Visibility Foggy Environment

Foggy weather poses significant challenges to traffic safety, as it reduces visibility and thereby increases the risk of accidents. This paper proposes a real-time vehicle detection and collision avoidance system for driver assistance that operates in fog, built using the deep learning YOLO-V8 architecture. It utilizes the lightweight YOLO-V8n and YOLO-V8s models. A new dataset of 1,900 images of Pakistani roads in foggy weather was collected and labeled for six types of vehicles, including rickshaws. The vehicle detection performance was evaluated using three datasets: Self-Made, Vehicle Detection in Adverse Weather Nature (DAWN), and Foggy Cityscapes (FC). YOLOV8n outperformed with a mAP50 score of 85% and 73.7 % on the self-made and DAWN datasets, respectively, compared to YOLO-V5s. Furthermore, YOLO-V8s outperformed with a mAP50 score of 65.1% on the FC dataset compared to YOLOV8n. The FPS rate of YOLO-V8n is 106, 114, and 135 for self-made, DAWN, and FC datasets, respectively. The TensorFlow Lite library is used to deploy a real-time vehicle detection system within an Android application, ensuring fast and efficient on-device processing. It provides warnings through visual, audio, and haptic feedback, making it a convenient way to enhance road safety in low-visibility conditions.