Research And Implementation Of Human Pose Estimation System Based On Convolutional Neural Network

Human pose estimation is the key technology to realize motion recognition and humancomputer interaction.In the multi-person scene,the position information of each keypoint of the human body needs to be detected at the same time.The flexible convolutional neural network architecture significantly improves the accuracy of this detection task.In this paper,we focus on two key factors of large convolution kernel and multi-scale fusion in network architecture.this paper proposes a multi-scale feature fusion of multi-stage cascaded convolutional neural network architecture,The first stage can roughly estimate the position of human body parts and the association degree between them in the small area,subsequent stages can be further refinement pose estimation results in the larger area.This paper mainly completed the following work:(1)Aiming at the scene of multi-person pose estimation,a multi-scale feature fusion of multistage cascaded convolutional neural network architecture is proposed,and the real keypoints detection confidence map is constructed by means of the heatmap regression coordinate.PAF method is used to construct the real keypoints association confidence map and complete the training of the network model.(2)Based on the completed training model,predict the detection confidence map and association confidence map of keypoints for the image after data enhancement,visualize the prediction result and test the learning effect of the model,and evaluate the accuracy of the model on COCO2017 validation set.In the experiment,four different scales of images were used,the body's keypoint detection confidence map and the association confidence map are obtained by keypoints detector and associated apparatus respectively,visualizing the degree of confidence and the degree of association between parts is to observe that the same human body response values in different images are different,so in the test we should increase the diversity of data to improve the detection effect of the model.Finally,the similarity between the predicted keypoints and the real keypoints was calculated by using the object keypoint similarity method,so as to obtain the average accuracy and average recall rate of the human pose estimation model based on multi-scale feature fusion.Compared with the original model,the large human body detection accuracy in the picture was improved from 12% to 16%.(3)Based on the detection confidence map and association confidence map of human keypoints,the coordinate of each part of human body and the association degree between different parts were obtained,and finally the multi-person analysis was completed with the help of the Hungarian algorithm,and the complete multi-person pose estimation results were obtained.Based on the above research results,a human pose estimation system is designed and implemented.