Due to the limitations of heated infrared imaging mode,infrared images in traffic scenes suffer from low contrast,diversity of target dimensions and postures,and mutual occlusion between targets,which leads to the decrease of detection accuracy and the false detection of some targets.In this paper,an improved algorithm based on YOLOv5s is proposed in terms of data processing,the AHE algorithm is used for partial data enhancement of the training set images；in terms of model improvement,YOLOv5s is improved by introducing a cross domain transfer learning strategy,inserting channel attention mechanism SENet and improving the loss function GIoU to α CIoU.By means of ablation experiments,the driving behavior of electric bicycle in night road environment is detected on self made data set.The experimental results show that the improved algorithm achieves an average accuracy of 95.9% for single person driving electric bicycle behavior detection,which is 3.1% higher than that of YOLOv5s；an average accuracy of 88.4% for manned electric bicycles behavior detection,which is 9.5% higher than the detection accuracy of YOLOv5s；and an average accuracy of 92.2% for total category detection,which is 6.4% higher than the detection accuracy of YOLOv5s,effectively reducing the probability of missing and false detection of infrared targets.