Research On Model Optimization Method Of Lightweight Human Pose Estimation

Human pose estimation is the basis of the algorithm of action recognition,behavior analysis,human-computer interaction and so on.With the popularity of intelligent devices such as robot,surveillance camera and smart home,human pose estimation algorithm needs to be integrated with other algorithms.At present,most of the algorithms of human pose estimation have a large amount of computation,and the speed of operation is slow in the equipment with limited computing capacity.Therefore,the optimization method of lightweight human pose estimation model is an important research direction.Firstly,the development history of human pose estimation algorithm is briefly reviewed.Then,the lightweight model is investigated.The optimization method of lightweight human pose estimation model is divided into three aspects: feature extraction network,joint point detection and joint point connection,The model optimization method with high cost performance is obtained,which is the balance between time and hardware cost and optimization effect.Based on the research of the current optimization methods of lightweight human pose estimation model at home and abroad,this paper adopts some model optimization methods with high cost performance to optimize the lightweight human pose estimation model,in order to achieve better performance.The main contents of this paper are as followsFirstly,aiming at the thickness of joint points and limbs,a two-stage lightweight human pose estimation network based on object detection is proposed.Through the idea of object detection,the pixel level detection of joint points is optimized to grid level detection,which makes the calculation of the model greatly reduced.At the same time,the performance of the model network structure is optimized according to the analysis of the thickness of joint points and limbs in the image.Then,this paper proposes a lightweight joint point layered detection method based on object detection.The two-stage human pose estimation algorithm is transformed into a single stage lightweight human pose estimation model,which improves the detection speed of the whole model.At the same time,the problem of two-stage detection model optimization can be transformed into one-stage detection model optimization,and network optimization can be more focused.Finally,a new joint detection model based on edge detection technology is proposed.Aiming at the accuracy and learning cost of lightweight human pose estimation model,the edge contour features extracted by edge detection technology are used to integrate with the features of network,which can help and correct the recognition of human posture.Among them,the method of foreground segmentation feature fusion is used to assist root node detection,and edge feature fusion is used to assist joint point detection..