Research On Human Pose Estimation Based On Lightweight Convolutional Neural Network

Human pose estimation is the process of obtaining key point information of the human body from images and constructing a human skeleton framework.In the fields of humancomputer interaction,behavior detection,virtual reality,etc.,it has extensive basic application and research value.The attitude estimation method represented by the convolutional neural network is the mainstream at present.But there are still many problems in its development:(1)The convolutional neural network has obvious shortcomings in the detection performance of complex and diverse human postures,and there is still a lot of space for improvement.(2)The human body pose estimation algorithm based on convolutional neural network has the problems of large amount of calculation and high memory consumption.(3)In complex scenes with crowding,occlusion,and insufficient lighting,the anti-interference ability of the pose estimation algorithm is weak.The main work of this thesis is as follows:Aiming at the computationally intensive problem of existing pose estimation algorithms,this thesis proposes a lightweight method for multi-way feature attention fusion.The model is based on the Higher HRNet network for lightweight design and training,including adopting channel splitting and channel shuffling to solve the information isolation between feature layers after group convolution;adopting the feature generation method of linear operation to solve the redundancy between different feature layers;adopting the method of fusing attention information to alleviate the accuracy drop caused by lightweight.The training,testing,visualization and ablation experiments of the model were completed on the MS COCO dataset.The experimental results show that the lightweight method in this thesis can significantly reduce the calculation amount of human posture estimation task under the premise of ensuring intuitive detection accuracy.In complex and changeable human body poses and complex scenes,human body pose estimation algorithms have the problems of low detection accuracy and poor anti-interference ability.Based on the lightweight network,this thesis proposes a human pose estimation method based on convolution fusion self-attention.The method includes Stem module responsible for splitting an image into image blocks,a Transformer module for extracting global features based on a self-attention mechanism,a lightweight convolution module for extracting local features,and a feature fusion module responsible for fusing global features and local features.This method completes model training in multiple data sets,improves the generalization ability of the model,and groups the difficulty level of key point detection for training,and improves the anti-interference ability of the model in complex scenes.The experimental results show that the method reaches the optimal level of the same level of lightweight methods with an average accuracy rate of 69.4%,which verifies the feasibility of completing the task of human pose estimation based on global features.