Research On Gait Recognition Based On Human Silhouette Features

In recent years, there has been an increased tendency of terrorist attacks occurring frequently, which makes people clearly understand the importance of human identification in national defense and public safety. Many biometric technologies(e.g., face, fingerprint and gait) have emerged for identifying and verifying individuals. Compared to other biometric methods, gait recognition offers several unique characteristics, such as high collectability and acceptability and low imitability, which makes it becomes the sole biometric technique that can be used for identifying people from a far distance. Therefore, gait recognition has become the hotspot of academic and industry fields. Recourse on gait recognition can play a very importance role in promoting the development of domestic intelligent monitoring system and maintaining social stability. We proposed three different gait features on the base of different human silhouettes features. All the three gait features can be used for individual identification. The main work of this paper includes:(1) Two new gait period detection algorithms are proposed. The main hypothesis of those two methods is that human would swing their arms and legs during their walk. Experimental results on a dataset show that our period detection metrics can achieve the best performance. At the same time, to our best knowledge, we are the first to use a quantitative method to assess the gait cycle estimation metrics. We utilize manually tagged information to assess the accuracy of different cycle detection methods. It has been shown that the second algorithm we proposed can achieve the best result.(2) We proposed three different gait features. Based on the theory of dense sampling, we put forward three different gait features, namely Frame difference HOG based gait feature(FHOG), gait feature based on Histogram of flow(GHOF), and Motion Boundary Histograms-based gait feature(MBH). Those three features are the static and dynamic features of human silhouettes. Extensive experiments show that the three new gait features all could acquire good recognition results. Furthermore, a novel approach to generate synthetic gait templates is proposed to address the problem of lack of gallery gait features. The recognition result on the synthetic templates even exhibits superior performance in comparison with that of the true templates.(3) A novel dimensionality reduction method is utilized in the procedure of feature classification and recognition. We combine the traditional PCA and LDA with NMF to process the high dimensional gait features. We also use information fusion to promote the overall gait recognition accuracy. Different gait feature are fused on decision level using a Sum-rule to improve the performance of identification. Experimental results on USF database show that our methods can achieve 67% accuracy.