Improved YOLOv5s Network for Traffic Object Detection with Complex Road Scenes

Aiming at the problems of low target detection accuracy and prone to false detection and missed detection in complex traffic scenes, a road target detection algorithm based on improved YOLOv5s is proposed. Firstly, a feature extraction layer based on Spatial Pyramid Pooling(SPP) is adopted for increasing shallow feature receptive fields named as Multi-scale SPP module (M-SPP). M-SPP is designed to fully integrate feature information of different scales for improving feature extraction ability of our network. Secondly, a Global Context Attention Module(GCAM) is proposed, in focusing on important features in channel and spatial dimensions, allowing the network to focus on more useful information. Finally, experimental results show that detection accuracy of the improved YOLOv5s model we proposed on the UA-DETRAC dataset, compared with the original algorithm, improved by 1.83% and 2.54% respectively. Our improved model abtain good detection accuracy while ensuring real-time performance, and solves problem of missed detection and false detection of small targets in complex road scenes to a certain extent.