Impact of the preprocessing stage on the performance of offline automatic vehicle counting using YOLO

Abstract

Vehicle counting systems detect, classify, and count vehicles with sensors or image processing, providing valuable information for road management. Image processing systems provide detailed information on vehicle flow with adequate lighting conditions and a higher computational cost compared to sensor systems. The image processing systems with higher accuracy require higher computational cost. This feature limits the number of application cases in cities with low technology level. This research analyzes urban vehicle counting using an automatic image processing system using YOLOv5 in the vehicle detection-classification stage and the SORT algorithm in the tracking stage. The study used videos recorded from a pedestrian bridge in Popayan, Colombia, for an exploratory study of the influence of preprocessing operations on the performance of a low-tech vehicle counting system. The study performed a comparative statistical analysis to determine the impact of different settings on system performance. An ANOVA analysis evaluates the incidence of frame cut and reshape on YOLO processing. The results indicate that a 30% cut of the image area prior to YOLO processing produces the lowest weighted average error. In addition, the frame reshape only increases the processing time. The study proposes improvements in the performance of an offline automatic vehicle counting system from the video preprocessing stage.