Gait-based person identification using multi-view sub-vector quantisation technique

This paper presents a novel method of person identification based on multi-view gait analysis. Three different views have been considered: side-view, front-right and rear right. The feature vectors set are derived from the sequence of outer contour width of binarised silhouettes of the walking person. The feature vectors of each view are segmented into sub-vectors and then quantised independently using Linde-Buzo-Gray (LBG) technique. This represents the gait signature using low dimensional vectors. Dynamic Time Warping (DTW) is used for gait signature sequence matching. Matching cost fusion technique is applied at the decision level to improve recognition performance. Experimental results demonstrate that decision fusion strategy of multi-views produces promising results compared to individual view based identification. Carnegie Mellon University's (CMU) dataset was used to test our algorithm. We achieved a recognition rate of 100% for slow walk and 92% for one walk-cycle of fast walk matching using multi-view decision fusion strategy for sub-vectors set size of 8x16.