Multi-Task Learning of Confounding Factors in Pose-Based Gait Recognition

This paper proposes a method for performing gait-recognition using skeletons extracted from human pose-estimation networks. Gait is a powerful biometric feature that has been used successfully to identify people, even in the presence of confounding factors such as different view angles and carrying/clothing variations. While most methods make use of Gait Energy Images (GEIs), we propose MFINet, a novel method for processing a sequence of skeletons extracted from an available pre-trained human pose estimation network, that incorporates confounding factors in the decision process. Inspired by methods in the area of activity recognition, we used a skeleton image representation (TSSI) in our experiments to capture temporal dynamics, as well as the skeleton spatial structure. Based on an extensive evaluation on the popular gait-recognition CASIA-B dataset, we show that MFINet is performing better than existing state-of-the-art pose-based methods, obtaining an accuracy of over 85% in scenarios with the same angle for both gallery and probe sets.