Yoga Poses Recognition Based On Mask R-CNN

Object detection and recognition using computer vision has been a very interesting and a challenging field of study from past three decades.Over the past decade or so,the rapid improvement of computing power has made breakthrough progress in deep learning,re-stimulating the interest of researchers in this field.Classification and target positioning based on machine learning and computer vision technology have always been popular in research,and have made great achievements.The pursuit of physical fitness and sports techniques by professional athletes has never ceased to be explored.People of all ages pay more attention to sports.Professional athletes can assist training through automatic or semi-automatic systems to improve training effectiveness,while ordinary people who love sports can use self-study and corrective training through appropriate systems.This demand has prompted researchers to combine artificial intelligence with sports to conduct research.This article focuses on the study of yoga recognition in the field of fitness.According to the task requirements of target detection,the classical convolutional neural network for common feature extraction and classification is analyzed.Considering that the deepening of the network can easily cause the gradient to disappear,this paper uses the deep residual network as the backbone network for identification and detection.The deep residual network is formed by cascading multiple residual blocks.The improved residual network adds a batch normalization layer on the basis of the classic residual unit to improve the network learning ability and performance.And use the face key point detection data set to verify the deep residual network.Based on the image task of yoga motion recognition,a yoga motion recognition method based on Mask R-CNN is proposed.The improved Mask R-CNN model is based on the framework and structure of the regional convolutional network.A certain number of candidate regions are proposed and classified for the image through feature extraction,and then these regions are output as detected bounding boxes,and the candidate The region uses segmentation branches for mask prediction.The improved Mask R-CNN model uses the improved deep residual network as the feature extraction backbone network,uses ROIAlign to bilinearly interpolate the extractedcandidate regions,then performs target classification and detection,and uses segmentation branches to segment the image.The model improves the convolution part of the split branch,replaces the original standard convolution with the depth separable convolution,and improves the network efficiency.The experiment builds a polygon-labeled data set and uses algorithms for simulation.The deepening of the network and the use of deep separable networks improve the accuracy of detection on the basis of maintaining network reliability,and verify the effectiveness of improving the Mask R-CNN network.