A lightweight vehicle detection algorithm combining image enhancement and multi-scale feature fusion

Abstract

In order to solve the problem of the high rate of missed vehicle detection in night-time traffic scenarios due to insufficient illumination and variable light sources, and the high complexity of the algorithm that makes it not well suited to in-vehicle devices. In this paper, we propose an efficient and accurate night-time vehicle detection method. Firstly, an improved EnlightenGAN algorithm is proposed to enhance the vehicle features in nighttime road images. Secondly, a lightweight network MobileNet v3 is proposed to replace the original backbone network Darknet53 of YOLO v3 for feature extraction as well as a multi-scale feature fusion strategy to improve the feature extraction capability of the network. Experimental results on the ExDARK dataset show that the accuracy and recall of the detection algorithm proposed in this paper are 88.88% and 94.32%, which are 0.53% and 8% better than YOLO v3, respectively, and the computational effort is reduced by 75% compared to YOLO v3.