Comparative Analysis of Single-Stage YOLO Algorithms for Vehicle Detection Under Extreme Weather Conditions

Abstract

Efficiency of vehicle detection algorithms in Computer Vision is predominantly calculated under clear weather condition; good weather with favorable lighting. Benchmarking of top-performing algorithms still reports a relatively low performance for vehicle detection under extreme circumstances. This is because vision-based detection algorithms face challenges in dealing with low quality images with background noise, bad lighting, and weather-caused distortions. The literature has reported substantial work on restoration of images prior to object detection. Nonetheless, the measures affecting the vision based algorithms and their effectiveness as well as how much degradation of the quality of the input image reduces the detectors output are less investigated. This paper focuses primarily on single-shot detector, which is You Look Once (YOLO) algorithm along its variations to detect vehicles real-time. The comparative experiments are set to evaluate their efficiencies based on precision, mean average precision (mAP), and recall rate using distorted vehicle images from the AAU Rain Snow Dataset. Based on our experimental analysis, YOLOv3 performed better as compared to other variants of YOLO. These findings will be used as the benchmarking results for improvement of vehicle detection algorithms under extreme weather conditions.