Research And Application Of Human Pose Estimation Model Based On Multi-scale Feature Learning

In the security industry,it is necessary to conduct real-time analysis of human behavior in surveillance video to avoid the occurrence of security accidents,and the current mainstream human behavior recognition methods are to judge by key points of the human body.It can be seen that using the information of human key points for behavior analysis has important research significance and practical application value.However,in the human keypoints detection model,there are problems such as long inference time and low inference accuracy caused by poor multi-scale feature learning ability.In this regard,combined with the practical application requirements of a chemical plant,this subject has carried out in-depth research on the human body pose estimation model.Firstly,a human pose estimation model is designed on the base of multi-scale feature learning.The model adds a multi-scale feature learning module on the base of the Trans Pose model,performs down-sampling and up-sampling operations on the input feature map to obtain feature maps of different scales,and then maps them to the same two-dimensional space as the input feature map.Fusion is performed to obtain the output feature map of the multi-scale feature learning module.This module improves the multi-scale learning ability of the model and enhances the semantic and spatial information of the feature map without affecting the size of the input feature map.Then,the model is trained and tested on the public dataset COCO.The human annotation information in the dataset is analyzed,and suitable data is extracted for training and testing,and the model is compared with similar models in terms of detection accuracy,inference speed and model size.The experimental results verify that the detection performance of the multi-scale feature learning model is better than other models.Finally,aiming at the actual needs of a chemical factory,a human behavior detection system for hand-held stairs is designed and implemented with the human pose estimation model based on multi-scale feature learning as the core.The system includes functions such as human detection,pose estimation,hand holding the stairs detection and result visualization.From the test and trial operation,the system runs stably and meets the requirements of users.