PERFORMANCE ANALYSIS OF YOLOV4-TINY-BASED FORWARD COLLISION WARNING SYSTEM ON RASPBERRY PI 4

Abstract

Traffic accidents caused by insufficient driver reaction time continue to be a major concern in road transportation safety. Forward Collision Warning (FCW) systems are widely recognized as an effective solution for reducing collision risks by providing early alerts to drivers. This study presents the development and performance evaluation of a real-time FCW system based on computer vision using YOLOv4-Tiny deployed on Raspberry Pi 4. The object detection process was implemented through the OpenCV Deep Neural Network (DNN) framework utilizing a pretrained YOLOv4-Tiny model trained on the COCO dataset without additional retraining. Experimental evaluation was conducted using five self-recorded driving videos under daytime, moderate-light, and nighttime environments. The system was configured to detect four vehicle categories, namely cars, motorcycles, buses, and trucks. Distance estimation was obtained through an inverse relationship between object size and bounding box width, while warning activation was triggered when the estimated distance was below 140 cm. Experimental findings demonstrated an average detection accuracy of 88.1%, processing speed ranging from 14 to 19 FPS, and an average response time of 0.35 seconds.