Electric Bicycle Detection Based on Deep Learning

Abstract

In China's urban traffic, the number of electric bicycles is increasing. Therefore, it becomes particularly important to accurately detect the behavior of electric bicycles and their riders through road traffic monitoring and implement efficient supervision to provide technical support. In the actual traffic surveillance video, electric bicycles occupy a small video image area and are easy to block each other, resulting in inaccurate detection and missed detection. To solve these problems, based on the idea of YOLOv4 algorithm, an improved detection algorithm of electric bicycle is proposed in this paper: replace the original YOLOv4 backbone network CSPDarknet-53 with GhostNet to enhance the detection speed. ECA attention mechanism is introduced in front of the three-layer prediction network to enhance the detection accuracy. The SPP module is replaced by the enhanced receptive field RFB module to strengthen the feature extraction ability. The experimental results show that the detection accuracy of the improved YOLOv4 algorithm is increased by 1.53%, and the detection speed is increased by 14FPS.