A High-Precision Fast Smoky Vehicle Detection Method Based on Improved Yolov5 Network

Abstract

A high-precision and fast smoky vehicle detection method was proposed. Since the existing target detection models deployed on embedded devices cannot meet the needs of rapid detection, an improved lightweight network based on Yolov5 was adopted in this paper. The backbone of Yolov5s was improved by Mobilenetv3-small to reduce the number of model parameters and calculations. In order to detect motor vehicle exhaust with high precision, a vehicle exhaust dataset is collected and established. Due to the interference of vehicle shadows and the occlusion between vehicles, Cutout and saturation transformation were applied to expand the self-built dataset, which was finally expanded to 6102 images. Experiments results show that after using data augmentation, the detection accuracy is increased by 8.5%. The improved network is deployed on embedded devices, and the detection speed of the network can reach 12.5FPS, which is 2 times higher than Yolov5's. The amount of improved network parameters is only 0.48M. This research proposes an efficient target detection model, and provides a possible method for the development of low-cost and rapid vehicle exhaust detection equipment.