Research On Gait Recognition Based On Kinect Skeletal Spatiotemporal Features

Gait recognition is a new authentication technology based on the gait when people are walking.Kinect can realize the spatial positioning and tracking of human skeleton,and output the 3D skeleton information of gait motion in real time,providing effective skeleton data for gait recognition.Aiming at tackling the problems of data normalization and feature extraction which have not been well tackled so far,two kinds of human recognition methods based on Kinect 3D skeleton data are proposed in this thesis.One is a single-frame human recognition method based on joint coordinate image and another is gait recognition method based on gait spatiotemporal feature matrix,Furthermore,an authentication system is formulated in this thesis.The research work is mainly as follows:(1)A single-frame human recognition method based on joint coordinate image is proposed considering the gap in the current research that bunch of studies heavily relied on manual selection of static features.This method normalizes a single frame of 3D skeleton data firstly,and then uses the "大" structure to spatially layout joint coordinate pixels.The joint coordinate image not only save the relative original data of the single frame 3D skeleton data,but also retain the natural connection structure of human skeleton.The classifier based on convolutional neural network is used to perform feature extraction and image classification,leading to the realization of the self-adaptive extraction of static features based on the target-oriented human recognition and the human recognition based on joint coordinate image.Experiments show that this method has the best recognition effect in the 10 people authentication task,and can achieve a recognition accuracy of 96.86%.Compared with similar algorithms,this method has higher human recognition accuracy.(2)A gait recognition method based on gait spatiotemporal feature matrix is proposed,aiming at tackling the problem that the single-frame human recognition method based on joint coordinate image only used static skeleton features for human representation,which cannot achieve high recognition accuracy in largescale authentication.In this thesis,the gait trajectory map is used to periodically measure 3D skeletal gait data,and then the spatial features of the skeleton and the motion features of adjacent skeleton frames are extracted to form gait spatiotemporal feature matrix.The gait spatiotemporal feature matrix extracts feature from the gait cycle sequence,which can effectively characterize the gait law of pedestrians and reduce the interference of redundant skeleton frames.Then,a classifier based on convolutional neural network is used for gait recognition.Experiments show that the method achieves 95.13% recognition accuracy in the164 people authentication task.Compared with similar algorithms,this method has stronger authentication ability and higher robustness.(3)An authentication system based on Kinect V2 skeleton tracking is built.The system uses Kinect V2 as the data acquisition device,uses C# and Python language to develop the system platform,and realize the functions of 3D skeleton data acquisition,human registration,classification training and real-time authentication.The core algorithms of the system are the single-frame human recognition method based on joint coordinate image and the gait recognition method based on gait spatiotemporal feature matrix.The system can choose the type of algorithm according to the collected skeleton quality.By inviting multiple volunteers to test the system,the experiments show that the system has reliable authentication capabilities and can achieves an accuracy rate of 87.09%±3.51%.