Personal Identification Based on Joint Motions with Error Correction Using Particle Filter

Gait analysis is useful for personal identification in public spaces. Since the emergence of deep learning that can accurately estimate the joint locations of a human in an image, gait analysis has become feasible in practical applications that have only a visible light camera. This paper proposes a novel method for personal identification based on joint motions that can improve utility of gait analysis. Specifically, the proposed method adopts a particle filter to correct the misclassification of left and right legs using the time-series input of their locations. Experimental results using datasets comprising real-world images showed that the proposed error correction improved the accuracy of personal classification from $84.19 \%$ to $85.94 \%$ when joint locations were estimated using ViTPose.