Violation target detection based on video streaming

Abstract

Based on the increasing number of pedestrians and non-motor vehicles running red lights, the use of video stream to detect illegal targets, obtain evidence of violations, as a basis for punishment, can effectively reduce the occurrence of violations. Based on the YOLOv3 algorithm, pedestrian and non-motor vehicle detection can be obtained by combining skin color detection and face detection, and redundant target information can be filtered by location score function, which can reduce the misjudgment of pedestrians. For non-motor vehicle testing, the cyclist’s position is determined by the re-matching of face or pedestrian position with non-motor vehicle. The border regression operation is carried out on the prediction box to make the non-motor vehicle detection box contain the information of cyclists.