Accuracy of personal identification based on joint motions extracted from 2D positions of a reduced set of joints

摘要

Personal identification based on gait analysis is suitable for surveillance at outdoor scenes and public facilities since it works well when images used for analysis are captured from distant viewpoints. It can also provide sufficient information for consumer behavior analysis, which has become significant in recent big data analysis. We propose a scheme based on joint motions for gait analysis; however, there has not been sufficient evaluation of schemes that we previously proposed. In this paper, we discuss the sufficient feasibility of our scheme by evaluating its accuracy regarding personal identification when feature extraction is conducted with two-dimensional coordinates of a reduced set of joints. Experimental results show that the average accuracy was 75.44% when a target human was captured from directly in front view and 78.07% when the target was captured from right in side view.