2D Human Pose Estimation In In Strong Multi-Person Interaction Scenes Based On Deep Learning

As an important research branch in computer vision in recent years,human pose estimation has been widely concerned by many researchers.At the same time,human posture estimation technology has been widely used in many practical applications.For example,automatic driving,medical treatment,virtual reality and human-computer interaction,sports analysis and so on.However,when facing the complex multi-person strong interaction scene,the human pose estimation network often produces some problems,such as inaccurate estimation of the key points of the occlusion part,poor estimation results of the multi-scale human pose,and lack of appropriate data sets.This paper focuses on the human pose estimation which is based on deep learning method in multi-person strong interaction scene.On the basis of analyzing the characteristics of the multi-person interaction scene,the appropriate posture estimation network structure is selected from the perspective of the multi-scale human posture estimation and the prediction of the occlusion key points,The Interact-Pose dataset based on multi-person strong interaction scene is applied to improve the performance of multi-person strong interaction human posture estimation algorithm.The main work of this paper is divided into the following aspects:(1)In view of the problem of multi-scale human bodies in the scene of strong multi-person interaction,this paper proposes a network called multi-resolution representation module to solve this problem.The module combines the two sub-modules of the multi-resolution representation serial module and the multi-resolution representation parallel module,fuses the results of the two sub-modules to predict the thermal map of the key points of the human body,and obtains the output results of the key points of different human bodies and their connections through association embedding and deconvolution operations.A large number of comparative experiments for other algorithms and ablation experiments for different sub-modules in the multi-resolution characterization module have been carried out on MSCOCO2017 and Crowd Pose data sets.The experimental results show that the multi-resolution representation module proposed in this paper has improved the results of human posture estimation in the scene of strong interaction between multiple people,and also has significantly improved the detection accuracy of multi-scale human body.(2)Aiming at the occlusion problem of a large number of human key points widely existing in multi-person strong interaction scenes(including direct self-occlusion of the same person and mutual occlusion between different human bodies),this paper proposes a human pose estimation structure based on the attention mechanism.The core in this algorithm,attention mechanism module,includes a channel attention sub-module and a spatial attention sub-module.The former realizes the identification of the correlation degree of different key points,and the latter improves the performance of the global context of the network.Through the fusion of the feature images generated by the two sub-modules,a more accurate key point prediction image is output,and the output image of the attention module is put into a multi-resolution parallel module for convolution operation to obtain the final key point prediction results.The experimental results show that the human pose estimation algorithm based on the attention mechanism proposed in this paper can effectively deal with the human pose estimation problem in the multi-person strong interaction scene,and has better performance than the current advanced human pose estimation algorithm on different datasets.(3)In view of a large number of occlusion problems of human key points widely existing in multi-person strong interaction scenes,the current common public data sets(such as MSCOCO,Crowd Pose,etc.)lack sufficient data in line with human motion requirements and scene requirements.However,it is difficult for algorithms lacking qualified training data to train models that can perform well in multi-person interaction scenarios through open data sets.Based on the laboratory multi-angle acquisition system,this paper collected a set of data including boxing,wrestling and other real actions that meet the requirements of the multi-person strong interaction scene research,and then enriched the background of the original data through the data enhancement scheme,and finally annotated the human body key points according to the annotation format of the public data set.The experiments on the human pose estimation algorithm show that the Interact-Pose data set can effectively improve the detection accuracy of human pose estimation in the multi-person strong interaction scene.