Research on Lightweight Dense Pedestrian Detection Algorithm
For the current problem of dense pedestrian detection tasks with many small size targets and high den-sity,low detection accuracy,a large number of parameters,and not easy to deploy,this paper proposes an improved lightweight dense pedestrian detection algorithm YOLO-GB based on the YOLOv5 algorithm.The Ghost module is in-troduced to form a lightweight backbone network,reduce the number of parameters and calculations,and extract image features at low cost.For the problem of a large variety of target scales,a prediction head is added to detect targets of different scales.A weighted bidirectional feature pyramid network BiFPN is introduced to enhance feature fusion and improve the multi-scale feature detection accuracy.Finally,we use Alpha-IoU to replace CIoU as the border regres-sion loss function to further optimize the detection accuracy.Experiments are conducted using the dense scene human detection dataset CrowdHuman,and the results show that the mAP50 of YOLO-GB reaches 84.8%,which is 1.5%higher than YOLOv5s,41.2%lower number of parameters,and 39.6%lower model size,with good detection accuracy and real-time performance.
NETL
NSTL
万方数据
