Researches On Multi-person Human Pose Estimation In Natural Scene

The development of computer vision has attracted much attention from industry and academia so far.This makes the machine imitate human vision as much as possible,so that it can automatically identify,analyze and perceive various activities of people and things in the surrounding environment.The law of human activity has become the first concern and research content.As one of the important research hotspots,human pose estimation has the main task of detecting key parts and joints of the human body in a given image or video,and outputting all or local body information.In recent years,thanks to the introduction and application of convolutional neural networks,human pose estimation has made great progress.At present,most attitude estimation methods mainly focus on improving the performance of attitude estimation while neglecting efficiency.At the same time,the pose estimation task itself has many difficulties such as scale differences,complex backgrounds,crowded occlusion,pose differences,and labeling errors.This makes the pose estimation task difficult to land,especially human pose estimation for natural scenes.Based on the above background,this thesis focuses on multi-person pose estimation in natural scenes.By analyzing the current research status of human pose estimation at home and abroad,in view of the difficulties in pose estimation,attentionbased selective pose distillation method and distillation-based model lightweighting method are proposed,focusing on general pose distillation method and lightweight model design and implementation.(1)This thesis proposes an attention-based selective pose distillation method.Firstly,a knowledge network with strong expressiveness and large amounts of parameters and a lightweight basic network were designed.Then,the attention-based selective pose distillation was used to transfer the structured knowledge of the knowledge network,which effectively improved the performance of the lightweight model on some ambiguous and under-learning samples expression and identification.(2)This thesis proposes a pose distillation-based model reduction.With reference to the principle of sea salt concentration and improvement of knowledge transfer gaps between multiple models,an automatic refinement cascade network knowledge strategy is designed,and heuristic learning is performed between the same models,and the model is greatly reduced while ensuring less accuracy of the model loss Parameters and calculations.In order to verify the effectiveness of the proposed method,a complete experiment was performed on the public data sets MPII and LSP collected in natural scenes,and the experiment verified the feasibility and robustness of the method.