A generic codebook based approach for gait recognition

Gait refers to the walking style of a person and it has emerged as an important biometric feature for person identification. The gait recognition algorithms proposed in literature exploit various types of information from the gait video sequence, such as, the skeletal data, human body shape, and silhouettes; and use these features to recognize the individuals. This paper presents the proposal of using a generic codebook in gait recognition. The idea is built upon a novel gait representation which exploits the spatiotemporal motion characteristics of the individual for identification. In particular, we propose to use a set of sample gait sequences to construct a generic codebook and use it to build a gait signature for person identification. To this end, we chose synthetic gait sequences of CMU MoCap gait database due to its diversity in walking styles. A set of spatiotemporal features are extracted from these sequences to build a generic codebook. The motion descriptors of real gait sequences are encoded using this generic codebook and Fisher vector encoding; the classification is performed using support vector machine. An extensive evaluation of this novel proposal is carried out using five benchmark gait databases: NLPR, CMU MoBo, TUM GAID, CASIA-B, and CASISA-C. In all experiments, the generic codebook is used in feature encoding. The performance of the proposed algorithm is also compared with the state-of-the-art gait recognition techniques and the results show that the idea of using a generic codebook in gait recognition is practical and effective.