A Novel Human Identification Method by Gait using Dynamics of Feature Points and Local Shape Features

Abstract

Gait analysis has been recently evolving techniques by which we can identify individuals using their gait patterns. One of appearance-based methods based on GEI (Gait Energy Image) has been used for forensic purpose in Japan. As long as the condition of two footages is close to an ideal case, which means the two footages are the same view-angle, clothing, resolution, frame-rate and stable walking, this method has been very useful so far. However, if the condition becomes far from an ideal case, the identification accuracy has become dropped down, resulting in analysis impossible. Here, we construct a novel human identification method based on comparison of dynamic features, which takes advantage of features of both appearance-based method and model-based method. Feature points (resemble to joint-points in model-based method) and those local shape features are semi-automatically extracted from silhouette sequences, and then the matching probability of two footages is calculated by comparing the dynamics of extracted features. It is found that GEI-based method is more useful in cases of frontal view, low resolution and comparison of multi view-angles, whereas the proposed method is more useful in cases of lateral view, low frame-rate and clothing variation condition. The results suggested that GEI-based method is superior to characterizing ‘figure’ information, whereas the proposed method is superior to characterizing ‘dynamic’ information of human gait.