| Human behavior understanding is an important branch of computer vision,which is widely used in public places such as sports,industry and schools.However,in complex public spaces,there are issues such as low algorithm accuracy and poor realtime performance when it comes to "human-object" and "human-human" interaction behavior.In order to improve the performance of human interaction behavior and achieve more accurate and faster human behavior understanding algorithm.In this thesis,we research the interactive behavior understanding method of pose reconstruction based on skeleton and image features.The research content of this thesis mainly includes: a lightweight skeleton extraction method based on improved OpenPose,a 3D human pose estimation algorithm under occlusion,and an interactive behavior understanding method of pose reconstruction based on skeleton and image features.The main contributions of this thesis are as follows:(1)In order to solve the issues of high model complexity and poor real-time performance in human pose estimation algorithms,this thesis proposes a lightweight human skeleton key-points extraction method based on improved OpenPose.First,build a multi-branch residual structure in the model training stage to speed up model training;Then,in the model inference stage,decouple the parsing output through reparameterization,and introduce HAM lightweight attention to further optimize feature extraction;Finally,experimental verification is carried out on the COCO human pose estimation datasets.The results show that,compared with the original model,the lightweight human skeleton key-points extraction method based on the improved OpenPose increases the inference speed by 47.3%,and improves the model precision by 4.3%.(2)Aiming at the problem of missing skeleton key-points in the output of human pose estimation algorithm caused by image occlusion,as well as the lack of missing 3D spatial information in 2D human pose estimation,this thesis proposes a 3D human pose estimation algorithm under occlusion.Firstly,on the basis of integrating skeleton correlation matrix,prediction algorithm is established to achieve interpolation prediction of missing skeleton key-points;Then,based on the complete human skeleton key-points,the nonlinear network is used to realize the 3D human pose estimation,and the OWM module is introduced to enhance the robustness of the model;Finally,experimental verification was carried out on the Human3.6M datasets.The experimental results show that compared with the maximum marginal neural network and other algorithms,the average prediction error of the 3D human pose estimation algorithm in this thesis is reduced by 13.8%,and the model has stronger skeleton extraction ability.(3)Aiming at the problem of low accuracy and poor real-time performance in interactive behavior understanding algorithm,this thesis proposes a method based on the improved OpenPose lightweight skeleton extraction method and the 3D human pose estimation algorithm under occlusion,named an interactive behavior understanding method of pose reconstruction based on skeletal and image feature fusion.This method fuses skeletal features and image features to improve the model’s environmental awareness and assist the model in understanding interactive behavior.Finally,through the experimental verification,the interactive behavior understanding algorithm of pose reconstruction based on skeleton and image feature fusion proposed in this thesis has increased by 6.1% in accuracy and 7.7% in speed,which shows that the algorithm in this thesis has the advantages of high real-time performance and high accuracy in interactive behavior understanding problems,it means that this algorithm has excellent application value.This article includes 50 pictures,19 tables,and 106 references. |
