Research Of Gait Recognition Based On Outermost Contour

With the increasing requirements for security, video surveillance in public places and facilities has become omnipresent. Human identification from the information provided by video surveillance is becoming an urgent requirement. Gait recognition, aiming to identify individuals by the way they walk, is a relatively new biometric identification technology. Compared with other biometric features, gait has its unique advantages. It does not require users’interaction and it is no-invasive. Also it is difficult to conceal or disguise. Furthermore, gait can be effective for recognition at a distance or at low resolution. Therefore, gait is very suitable for intelligent video surveillance. Actually, gait recognition is a hot research area now.Currently, most of the gait recognition methods are model free ones. And it has some problems. The gait recognition methods based on silhouette have plenty of redundant data, while the methods based on contour are sensitive to noises. In order to eliminate the effect of contour noise, in this thesis, we propose a gait recognition method based on outermost contour. However, gait is a behavioral characteristic. In comparison with other biometric features, it has inherent disadvantages, such as low accuracy, weak robustness. In order to improve the recognition performance of video surveillance system, we use multiple biometric feature information and propose a hybrid fusion recognition method using face and gait.The main content of this thesis contains two parts, gait recognition method based on outermost contour, hybrid fusion recognition method using face and gait.Considering gait recognition methods based on contour has the problem of complex computation and sensitive to noise, we propose the gait recognition method based on outermost contour which can effectively avoid the effect of noise. Firstly, we normalize all the silhouettes to the same height. And then we extract silhouette feature using outermost contour (in each row of a normalized silhouette image, the most right pixel and the most left pixel on the contour belong to outermost contour). PCA training is adopted to reduce the dimensionality of silhouette feature and gait feature is computed in the new feature space. In order to effectively classify the extracted gait features, three methods—MDA, BPNN and SVM methods are used for classification. Experimental results show that the gait recognition method based on outermost contour can achieve very high accuracy. Gait recognition has inherent disadvantages—low accuracy and weak robustness. A gait recognition system couldn’t meet the accuracy requirement of a recognition system. Considering face and gait can be acquired by the same video surveillance, fusion of face and gait for recognition can improve the performance and eventually promote the development of intelligent video surveillance, which has great theoretical and practical value. Therefore, we propose a hybrid fusion recognition method using face and gait in this thesis. The fusion method includes two parts—serial part and parallel part. These two parts are executed in order. In serial part, face recognition algorithm and gait recognition algorithm are used in turn to identify the unknown sample. In order to avoid false acceptance, we set a strict threshold for each algorithm in this part. If one of the algorithms can identify the sample, the identification process is finished and the hybrid fusion terminate. Otherwise, the parallel part will be activated, which uses rank-level fusion for recognition. Experimental results show that the accuracy of hybrid fusion method can achieve100%in the heterogenous database of100people. This method balances accuracy and speed, which also provides new solution to multi-modal biometrics.The proposed gait recognition method can effectively eliminate the effect of noise in the foot area generated by the imperfect segmentation of silhouette, and eventually improve the performance. However, we only consider the scenario of normal walking. We will study the robustness of the proposed method with respect to view angles, clothes and walking styles. In addition, we have made progress in the fusion of face and gait. Based on the achievement, we will continue study human identification by fusion of face and gait in complex environment.