Research On Gait Recognition Method Based On Improved GCN

Since the end of the last century,the rapid development of Internet technology has brought a large number of new security problems.Biometrics technology can effectively solve security problems such as authority identification,so biometrics technology also shines.Commonly used biometric technologies include face,fingerprint and iris recognition,but these biometric technologies all require close interaction between the target object and the identification device and are easily forged.The gait recognition technology uses the motion features generated by the moving objects to perform biometric identification.The gait features have the advantages of long-distance identification and difficult forgery,and solves the shortcomings in the abovementioned biometric identification technology.Most of the existing gait recognition methods use frames or videos generated by traditional cameras for recognition.Such methods often require a lot of resources to segment the background information and gait information in the image,and even affect the final result of gait recognition.The emergence of the event camera solves this problem.The event camera only pays attention to the illumination changes in the shooting area,so it can naturally ignore the background information and reduce the workload of gait recognition.In addition,the high sampling frequency of the event camera allows it to capture richer gait information,better adapt to extreme lighting environments,and capture fast motion.Using the event stream data generated by the event camera,this thesis designs a noise removal algorithm and down-sampling algorithm that combines the characteristics of the gait movement to preprocess the data,and proposes an improved graph convolutional network model(IGCN)to complete the task of gait recognition.Based on the traditional graph convolutional network,this thesis adds a data connection module for event stream data and multiple residual-pooling network layers,and introduces the Top K algorithm to improve the traditional graph pooling layer,so that the network model proposed in this thesis can better adapt to event stream data to complete gait recognition tasks.The algorithm proposed in this thesis and the latest related algorithms are evaluated and compared using the gait data collected in the real environment and the CASIA-B gait data.The experimental results show that the accuracy of the method proposed in this thesis can reach 98% for gait recognition on the gait data set collected in the real environment.Compared with the latest unsupervised denoising techniques,the noise removal method proposed in this thesis improves the gait recognition accuracy by 2%.Compared with directly using the octree grid filtering algorithm for downsampling,the down-sampling method in this thesis can improve the accuracy by 4.5% with almost the same down-sampling effect.In addition,the average accuracy of the algorithm proposed in this thesis is also better than other algorithms on the CASIA-B dataset,and in the night environment,the algorithm proposed in this thesis also shows better accuracy than other algorithms,with strong performance robustness.