Research On Human Gait Recognition Methods Based On Image Sequences

With the development of modern society and the increase of people safety consciousness, human identity recognition is indispensable to the safety of more and more public places such as stations, airports, banks, administrations, communities and so on. Biometrics recognition, a kind of identity recognition technologies, can identify human identity according to physiology or activity characters and have the superiority over traditional recognition metods in accuracy and speed. Biometric features used currently include ear, hand geometry, fingerprint, palmprint, face, iris, voice, signature, etc. As a new technology of biometrics recognition, gait recognition has its predominance among other identity recognition technologies because it has the advantages of being noninvasive and requiring little about the quality of video. Now many researchers have paid growing attention to gait recognition. Based on its strong points and importance both in theoretical research and practical application, gait recognition will be further studied in this dissertation. The main contributions of this dissertation are summarized as follows:(1) Proposed gait recognition methods based on shape features of multi-regions. Firstly, every human silhouette image is devided into several sub-regions in the video sequence. There are three kinds of sub-regions as follows: the whole silhouette image regarded as one sub-region, sub-regions with same sizes, sub-regions with different sizes that can be obtained according to properties of body segments extracted by using anatomical knowledge. Secondly, by means of extracting the silhouette or contour shape features of each sub-region and computing their changing features in the gait sequence, gait feature vectors can be constructed. Finally, experiments show that the proposed methods are valid and have good recognition results.(2) Proposed a gait recognition method based Radon transform. The Radon transform belonging to image transform can measure the projections of image in certain directions. The Radon transform can represent gait angular features. The reason is that during human walking, there is large variation in the angles formed by leg and arm swinging. The proposed method constructs a Radon transform template for every gait cycle in the gait sequence and extracts gait features from this template. Experimental results show that the proposed method can gain better recognition performance.(3) Proposed a subtractive clustering method based on genetic algorithms and a gait recognition approach based on main walking postures. The traditional subtractive clustering is modified and the genetic algorithms are employed to optimize the relative parameters in the improved subtractive clustering. Clustering experimental results show that the new clustering method can get the higher clustering accuracies than the traditional subtractive clustering. The new gait recognition approach firstly divides many walking postures of every sequence into several clusters by the subtractive clustering method based on genetic algorithms. Secondly, all walking postures from a cluster are averaged in order to gain one main walking posture. Finally, gait recognition is done by matching main walking postures among imge sequences. Gait recognition experimental results show that the new gait recognition approach can get the satisfying performance.(4) Proposed gait recognition methods based on energy images. If the average image is calculated after several images are added, it has little noise. Because energy images are formed by applying image addition to image average, they have the advantage of little noise. The proposed method constructs several energy images for every gait sequence and extracts gait features from these images to do gait recognition. There are two sorts of energy images as follows: key frame energy images including key frame energy images with maximal contour width and key frame energy images with minimal contour width, standard deviation energy images including zero standard deviation energy images and non-zero standard deviation energy images. Experimental results show that the proposed methods can achieve encouraging gait recognition performance.(5) Proposed a sphere-structured support vector machines classification algorithm based on a new decision rule, a novel kernel-based fuzzy hyperspheres classification algorithm, and gait methods based on hyperspheres classification algorithms. Firstly, because of using a new decision rule, the new sphere-structured support vector machines can gain the higher classification accuracies than traditional sphere-structured support vector machines. Secondly, the kernel-based fuzzy hyperspheres classification algorithm not only attain the better classification accuracies and lower computational complexity than traditional sphere-structured support vector machines and the hyperspheres algorithms based on moving median centers. Finally, two proposed hyperspheres classification algorithms metioned above are applied to relative gait recognition methods in order to further improve the recognition performance.