A lightweight model for multi-traffic object detection based on deep learning under complex traffic conditions

Abstract

ABSTRACT The object detection is extremely important in autonomous driving environment awareness. Besides vehicle and pedestrian detection, traffic signs and lights are important objects. The paper presents how to achieve precise results in multi-traffic object detection while minimizing the model size. A deep learning network YOLOv5s-Ghost-SE-DW is proposed based on the YOLOv5s. The proposed network can detect all traffic objects including traffic signs and lights. First, the convolution layer is replaced by Ghost module to reduce the parameter and model size. Second, in order to improve accuracy and real-time performance, the attention mechanism SELayer is embedded to fuse more spatial features. Third, the DW convolution is used to extract features and further reduce the parameter number. The effect of different modules on the whole network is verified by ablation experiments. The YOLOv5s-Ghost-SE-DW yields a model size of 5.22MB while achieving 15.58 FPS real-time performance on CPU. The FPS increases by 27.5%.