Robust frontal gait recognition – merging viewpoints and depth ranges

This paper proposes a frontal gait recognition system using a single camera, which is robust to changes in clothing and carrying condition. User silhouettes are derived from 2D plus depth (2.5D) sequences, using background subtraction. Silhouettes are integrated into a 3D point cloud, corresponding to a marching in place (MIP) representation of the sequence of observed silhouettes. Features are then extracted from frontal, top and side viewpoints of the MIP. Additionally, this paper proposes the novel usage of multiple depth range segments of the frontal silhouette view, to better exploit some of the user distinctive motion information. The Histogram of Oriented Gradient (HOG) descriptor is applied to each of the considered views and to three depth range segments. Fusion of the resulting descriptors is tested at feature, score and decision levels. The proposed method is evaluated on the IST 2.5D frontal gait dataset, composed of 30 test subjects, walking under different clothing and carrying conditions, acquired on different days. Experimental results show that combining the proposed descriptors outperforms state of the art methods, achieving a recognition rate of 100% for the considered database.