Research on Intelligent Estimation Method of Human Moving Target Pose Based on Adaptive Attention Mechanism

Meishuang Ding,Jing Zhao,Xiaohui Yuan

doi:10.1155/2022/2141194

Meishuang Ding, Jing Zhao + Show 1 more

Open Access

https://doi.org/10.1155/2022/2141194

Copy DOI

Abstract

In daily physical education, posture performance is an important basis for making excellent results. This paper explores an intelligent method to estimate the target pose based on adaptive attention mechanism. First, the regional attention is iteratively generated from a global level to a local level based on the attention mechanism. Human decision-making patterns are imitated to evaluate the effectiveness of regional attention in real time. The level of attention mechanism is adaptively adjusted and focused layer by layer to achieve precise target detection and tracking. Second, with the target frame obtained from each frame, the pose estimation algorithm finds the key points of human body, enabling the human body pose optimization strategy to solve the crossover problem of the key points. Results of experiments on sports video images show that the proposed method has a higher accuracy in pose estimation than other algorithms and can help sportsmen adjust their training methods scientifically.

Highlights

Nowadays, sports results account for an increasing proportion of a student’s total academic results
SiamRPN [6] constructs a Region Proposal Network (RPN) structure based on twin network, where the template image and the search region use the exact same convolutional network to extract the feature map, the classification and regression are carried out through two independent network branches to determine the target location and size, and the external frames are optimized by regression results
Convolutional Network (FCN) [7] is the classical network structure in semantic segmentation, where all network layers use convolutional layers, and the original image size is recovered by upsampling layers, so that the output segmented image size is exactly the same as the input image size. ose target tracking algorithms are still restricted by appearance deformation, light fluctuation, fast

Summary

Introduction

Sports results account for an increasing proportion of a student’s total academic results. Erefore, in order to solve the problems of the traditional target tracking network construction model and imitate the human cognitive model, this paper explores an intelligent target pose estimation method based on the adaptive attention mechanism. If the threshold is satisfied, positioning of the target to be detected in the background image is considered reliable, and the trained deep network is stored as the first model into the detection model set of the attention mechanism; otherwise, the order of attention is adaptively adjusted to provide heuristic information for the positioning network. Adjust the clipped image to the size consistent with that of the input image, and start the second round of test; enter the deep neural network M1 again to extract the deep feature map; repeat the above process to establish the second attention mechanism detection model, or generate a new image to be detected to initiate the round of test until the test requirement is met or the number of focusing reaches the threshold. Using Resnet-18 model M0, Uj is given the fully connected

Objective evaluation function

Pose Estimation Model Based on Deep Learning

Experimental Results and Analysis

Conclusions