Towards Feature Fusion for Human Identification by Gait

In this paper, we propose a statistical gait feature fusion approach for human recognition by gait. First, we produce a gait period estimation function by converting the contour of silhouette in specific regions into an ID signal, and divide each silhouette sequence into cycles. With a novel shape descriptor while retaining translation, scale and rotation invariance, a statistical feature extraction method is used for learning gait features from individual frame and consecutive frames, respectively. Features learned from individual frame characterize human silhouette properties, and features learned from consecutive frames describe dynamic properties of human motion. Next, we employ Jeffrey divergence and dynamic time warping for measuring the similarity between test and reference sequences. To improve the recognition performance, a fusion rule on silhouette and dynamic gait features is developed. Experimental results show that recognition performance achieved by the proposed feature fusion approach is better than that achieved by individual silhouette or dynamic feature classification approaches, and better than existed methods.