Research on High-Position Surveillance Suspicious Target Tracking Based on Public Security
Aiming at the problem of poor recognition and tracking effect of long distance suspicious targets(super small targets)in public surveillance suspicious target detection and tracking,a fusion algorithm based on improved target recognition and target tracking is proposed to detect and track super small targets and multiple suspicious tar-gets.Firstly,a pluggable P_CBAM attention module was constructed to enhance the weight of YOLOv5s model on key feature channels.Tthen,a YOLOv5s ultra-small target detection prediction layer was added,and K-Mean clustering was carried out on detection anchor frames to reduce the loss of target ID.Next,a non-maximum suppression method was used to remove detection overlapping frames.Finally,associating and fusing the target appearance in the anchor box detected by YOLOv5s with the motion information,Y5s_D_S target tracking model was constructed.The simula-tion results of ablation experiments and comparative experiments show that the Y5s_D_S model has the best detection accuracy and precision in ultra-small target detection and tracking compared with other baseline models on the MOT20 high-level monitoring data set.Compared with the traditional Deep_SORT model,the accuracy and F1 value of Y5s_D_S model are improved by 15.3%and 0.059,respectively.The fusion algorithm proposed in this paper has high superiority and robustness for small target detection in multi-target high-position surveillance video.
NETL
NSTL
万方数据
