Human Pose Estimation Method Based On Multi-scale Features Of High Resolution Network

As a major research branch in the field of computer vision,human pose estimation has been widely used in industrial vision inspection,security monitoring,robot-aided design,medical-aided analysis,behavior recognition and film post-production.With the increasing popularity of artificial intelligence and convolutional neural networks,human pose estimation methods using convolutional neural networks as the main building unit of models have made significant progress in performance.However,the network model of existing methods still has more limitations.Because the object scale is smaller at varying object scales and medium and far distances under the multi-person scene,the human pose estimation results do not work well.Complex background or occlusion also creates low precision rates for hard joints localization.To solve these problems has become a key problem in the research of human pose estimation.This paper will present a detailed introduction around the human pose estimation problem and investigate application problems such as real-time pose recognition in living scenarios.(1)In view of the problems in the existing pose estimation methods,the network model proposed in this paper is based on the high-resolution network and the multi-person pose matching network.In this paper,a residual down-sampling module is proposed for down-sampling,which can make up for spatial information loss and make multi-scale information fusion better with high and low resolution features.A dual attention mechanism based network learning guidance strategy,combined with down-sampling residual modules to learn more useful feature information,was simultaneously used to effectively improve the localization precision of hard joints.In addition,the MRDPose proposed in this paper is tested on the open source standard dataset.In the results of the MPII dataset,the total mean average precision reached 81.1%,and average precision of hard joints reached 74.4%.Meanwhile,the mean average precision reached 70.5% on the MS COCO dataset.The average precision of medium objects reached 67.7%.(2)Aiming at the existing security monitoring equipment,an intelligent improvement scheme is proposed in this paper.This paper presents a real-time pose recognition system,in which the system program applies the human pose estimation algorithm and the action classification algorithm based on the sequence of bone keypoints.It is trained and validated on the self-made abnormal pose dataset.According to the real-time requirements of the monitoring process,the model is improved.The theft activity(abnormal pose)recognition in the actual monitoring scene is realized.Finally,the functions of the whole system are visually validated and the experimental data are analyzed,which meets the requirements.