SVM And DSW Based Gait Recognition By Column Mass Vector

Gait recognition, recognizing persons through the way they walk, as an unique biometric identification technology, has it’s powerful potential to identify people at a distance. However,current research still trapped in theoretic exploring process, far from actual application. This paper mainly focuses on feature extraction, classification and identification processes. Based on existing gait recognition algorithms, a novel gait feature—column mass vector—is proposed for recognition. The main work and results are presented as follows:Feature of column mass vector is proposed to represent gait information, Column mass vector divides silhouette of person vertically, downsizes pixel-level computation complexity while preserving structure of walking people, and reflects coordinated moving of the body during walking process through periodical changes of the vector. Adopting Dynamic Spatial Warping(DSW)method to warp column mass vectors of probe set and gallery set, on the one hand, addresses translation problem resulted from extraction of moving object, on the other hand,improves recognition accuracy against interference factor especially carryings by distinguish interference area resulted from wearings and carryings out. In the paper, Dynamic Time Warping method is extended to Dynamic Spatial Warping, turning the matching process based on time period into space and taking full advantage of periodical characteristics of gait, decreases the vulnerability of gait period characteristic against phase shift.Furthermore, Active Energy Image(AEI) is utilized for dimension reduction. Via adding up the differential part between each neighboring frames of a period, we can avoid “dimension disaster” and reduce complexity of feature extraction drastically.Due to limited training set of gallery, the work adopts support vector machine(SVM)method which is based on small sample statistic theory, because SVM itself has it’s special advantage in solving small sample, high generalization, high dimension pattern problems. By applying one to the rest principle to identify different probes, we get promising performance in SVM and competitive recognition accuracy in DSW on the CASIA database under normal walking and thick clothing conditions, respectively.