Personal Identification Based on Joint Motions with Normalization against Size Changes of Targets

Gait analysis is useful for personal identification in public spaces. Recent, advancements in deep learning technology have enabled highly accurate estimation of human joint positions in images, making practical applications of gait analysis possible using only visible light cameras. We propose a method to improve the performance of personal identification methods based on joint movements captured by cameras, aiming to make them more practical for use in real-world scenarios. Specifically, we propose a novel normalization method for feature vectors to address variations in the size of target humans during walking sequences. Such variations have a significant impact on identification accuracy in gait analysis. The proposed method normalizes feature vectors based on the distance between the hip and neck joints, reducing the impact of differences in target human size in the images. To evaluate the effectiveness of this method, we created two datasets by intentionally varying the magnification of existing datasets. Experimental results showed that the proposed method improved identification accuracy by approximately $12 \%$. Furthermore, even when multiple datasets were combined, the accuracy increased from $75.5 \%$ to $84.7 \%$, demonstrating that the proposed method is highly robust under various conditions.