Multi-View Gait Based Human Identification System with Covariate Analysis

This paper presents a multi-view gait based human identification system. The system is able to perform well under different walking trajectories and various covariate factors such as apparel, loan carrying and speed of walking. Our approach first applies perspective correction to adjust silhouettes from an oblique view to side-view plane. Joint positions of hip, knees and ankles are then detected based on human body proportion. Next, static and dynamic gait features are extracted and smoothed by the Gaussian filter to mitigate the effect of outliers. Feature normalization and selection are subsequently applied before the classification process. The performance of the proposed system was evaluated on SOTON Covariate Database and SOTON Oblique Database from University of Southampton. It achieved 92.1% correct classification rates for both databases