Object Detection of Pedestrain and Vehicle at Night Based on Improved YOLO Algorithm

In the task of pedestrian and vehicle detection at night, challenges such as complex environments, small target sizes, and dense distributions exist. Aiming at these challeges, a novel YOLOv8-ECSH based on YOLOv8 is suggested in this research. Firstly, the lightweight up-sampling operator CARAFE is introduced. It aggregates information about the context with greater sense of wildness, thereby could enhance the algorithm's detection speed and accuracy. Secondly, the CIOU Loss is replaced by the EIOU Loss. This minimises the size of the difference in width and height between the predicted boxes and ground truth boxes, so that could enhance the speed of convergence of models. Thirdly, SPPCSPC module is inserted to backbone network to extend features in different scales of the target. Finally, one small target detection layer was added, while a large target detection layer is cropped out. It improves the detection of small targets by the model. The experimental results on a self-made dataset of pedestrian and vehicle at night show that the YOLOv8-ECSH model improves the recall by 4.1% and the mAP@0.5 by 1.8% to 92.6%.