基于轻量化DeepSort的人数统计

People counting based on lightweight DeepSort

为了检测学校教室或图书馆的人流量,设计了一种优化的YOLOv5和轻量化DeepSort的人流量统计系统。为提升目标检测效果,采用CIoU loss回归损失和DIoU-NMS非极大抑制,在加快目标边界框参数学习的同时可提高定位精度,提升遮挡行人的检测性能;在DeepSort外观特征提取网络的基础上,结合轻量级网络ShuffleNet V2,对特征提取模型重新训练,减小模型的参数网络复杂度并保持良好的精确度,提高了系统的可移植能力;在视频中设置虚拟线,利用行人通过虚拟线的时间差进一步计算行人速度。实验结果表明,采用端到端的方式对行人目标进行高效追踪,缩小后的模型体积仅为原模型的5%,大大改善了对遮挡行人的检测性能,可以较准确地统计出人流量与行速,并提高了鲁棒性。

Aimed at the flow of people in the classroom or library,an optimized YOLOv5 and a lightweight DeepSort flow statistics system are proposed.In order to enhance the target detection effects,the CIoU loss regression loss and DIoU-NMS non-maximum suppression are used to speed up the learning target bounding box parameters,improve the positioning accuracy and detection performance of occluded pedestrians.Based on DeepSort appearance feature extraction network and ShuffleNetV2 lightweight network,the feature extraction model is retrained to reduce the parameter network complexity and maintain good accuracy,and boost the system portability.Virtual lines are set up in the video to detect the flow of people,and the time difference between pedestrians passing the virtual lines is used to further calculate pedestrian speed.In this paper,the end-to-end method is used to effectively monitor the pedestrian target.Experimental results show that the proposed design represents only 5%of the original network volume,which improves the performance of occluded pedestrian detection,accurately counts the pedestrian flow and speed,and enhances the robustness.