Research On Attitude And Gait Recognition Algorithm Based On YOLOv3 Pruning Model

In the pattern recognition of computer vision,target detection is a very challenging and realistic research direction.As an important part of computer vision,multi-person pose estimation also affects many other research directions,including but not limited to human-computer interaction and somatosensory technology.With the trend of globalization becoming more and more obvious,the security problem is becoming more and more serious.Compared with other methods for disguising identity,gait has become the focus of national security research because of its non-disguise and longdistance characteristics.In recent years,with the development of computer hardware and software and the explosive growth of data volume,deep learning has performed very well in the field of computer vision.Multi-person pose estimation and gait recognition methods based on deep learning have also played an important role in various fields.The paper first introduces the research background and significance of multi-person attitude estimation and gait recognition technology,and elaborates the research status at home and abroad,including the mainstream gait recognition algorithm,and analyzes the problems existing in the current technology;With the great brilliance of convolutional neural networks and deep learning in the computer field,the paper carefully introduces the basic theoretical knowledge and important concepts of deep learning;This paper focuses on the following two aspects of work and innovation:(1)Firstly,a multi-person pose estimation algorithm YLPE(YOLO Pose Estimator)based on YOLOv3(You Only Look Once v3)is proposed.Based on the problem of excessive parameter size and computational redundancy,the model-based pruning is further proposed.Multi-person pose estimation algorithm YLPPE(YOLO Prune Pose Estimator).The algorithm adopts a top-down frame,and uses the YOLOv3 network to perform human target detection on multi-person pictures,and generates a new-sized single-person picture by cutting,zero-filling,etc.,and finally inputting singlepicture images sequentially.The Stacked Hourglass Network(SHN)performs human joint point detection and uses the central point regression method to return the detected joint points to the original picture.Through the experimental verification of the pruning model,the parameter amount of the YOLOv3 pruning model decreased by 46%,but the accuracy only decreased by 0.5%.The YLPE algorithm achieved 84.1m AP on the MPII dataset,while the YLPPE algorithm was 83.7m AP.At the same time,the algorithm is compared with the Deeper Cut model and the RMPE model to obtain a more accurate lead.(2)A gait recognition algorithm based on HP-GSI fusion is proposed.The current mainstream methods of gait recognition are based on Human Posture(HP)and Gait Silhouette Image(GSI).The former has problems such as inaccurate joint positioning,while the latter is worn by pedestrians and perspectives.Aiming at the above problems,this paper proposes a gait recognition algorithm based on HP-GSI fusion.The data set used by the algorithm is YLPPE algorithm for image size cropping and joint point information extraction from the original CASIA-B gait dataset.The algorithm is based on the HP gait recognition module and the GSI gait recognition module implemented by convolutional neural network.The module adopts the Set Pooling(SP)layer and the complete gait cycle picture feature to obtain the gait timing characteristics.The video is input into these two modules to obtain the recognition confidence and recognition result,and then the improved AND decision layer feature fusion method is used to obtain the final detection result.The pedestrian gait recognition algorithm based on HP-GSI fusion reached 75.8m AP on the CASIA-B dataset,which is improved compared with the individual identification module,which proves the effectiveness of feature fusion.