Gait Recognition from Freestyle Walks Using Relative Coordinates and Random Subsequence-Based Sum-Rule Classification

Because of the increase in terrorist attacks, human identification research has become more popular. Gait recognition is one of biometric recognition techniques that can be accurately used to identify a person, since it is hard to alter the style of movement continuously and permanently. In this work, we propose a new gait recognition technique for freestyle walks using Microsoft Kinect. Since our technique supports freestyle walks, it can be employed to identify a person from a distance. This technique can also be used to collect gait information non-invasively and without a person's awareness. Our proposed technique introduce two new concepts to cope with challenges that come with freestyle walking. First, relative coordinate concept is created to handle the non-fixed observation angle issue. Second, the random subsequence-based sum-rule classification is introduced to handle non-fixed length walks issue. Our proposed technique takes different local characteristics of a walk into consideration. It significantly outperforms baseline technique (Dynamic Time Warping with k-NN with accuracy rate of 92.22%. The results also suggests that a human recognition is better done from observing many small movements and postures than observing one large sequence of walk.