Design And Implementation Of Human Body Gesture Recognition System Based On Deep Learning And Robotics

Human body pose recognition is a hot research direction in the field of computer vision.Human body pose recognition is a process of detecting key points of the human body through images or videos.Through the key points of the human body,information about the posture of the human body can be obtained,so as to provide key information for understanding or imitating human behavior.In recent years,emerging fields such as gait recognition security systems,short video special effects,and bionic robots all need the help of accurate human posture information.Human posture recognition has a wide range of application prospects.With the continuous development of deep learning,the accuracy of human body pose recognition is getting higher and higher,which can already meet the accuracy requirements in most scenarios.However,in practical applications,models based on deep learning still have certain problems.Models based on deep learning usually have large parameters,high hardware requirements,and slow inference speed.In some scenarios that require high inference speed,such as real-time interactive robots,there is a speed bottleneck.At the same time,previous research work usually focuses on improving a single module,and there are few complete human pose recognition system solutions.Therefore,this paper designs a set of pluggable high-precision and low-latency human posture recognition system.The system uses video as input.After detecting human targets,it recognizes two-dimensional human body key points,converts two-dimensional human body key points into three-dimensional coordinates,and uses three-dimensional human body key points as output,which can be applied to scenes such as robot systems.This article uses semi-supervised learning and teacher annealing to improve the knowledge distillation method,and compresses the human body pose recognition model.Based on the HRNet model,this method can greatly improve the inference speed of the model with similar accuracy.Increase mAP by 1.2%and 0.5%respectively.On the COCO data set,the human body pose recognition model compressed in this paper reduces the amount of model parameters by about 67.4%and increases the inference speed by about 2.4 times when the accuracy is similar.At the same time,the overall system of human posture recognition constructed in this paper can recognize the coordinates of three-dimensional key points of the human body more accurately and efficiently,and provides a complete solution for human posture recognition,and it has good agility.