A Geometric View Transformation Model Using Free-Form Deformation for Cross-View Gait Recognition

Gait is a commonly used behavioral biometric that proved to be successful for authenticating people at a distance, however, its recognition accuracy severely deteriorates due to view changes. We therefore propose a Geometric View Transformation Model (GVTM) to enhance the robustness of gait recognition under cross-view conditions. Specifically, we train a subject-independent warping field with a free-form deformation framework which geometrically transforms gait features from two different views into those from an intermediate view. We then apply it to gait features of a test subject to register them and subsequently match them under the same intermediate view. Unlike existing appearance-based view transformation models that may corrupt the gait features, the proposed GVTM does not corrupt them because it preserves their spatial proximity. In addition, the GVTM can transform features more flexibly than simple weak perspective projection-based geometric approaches and more efficiently than 3D model-based approaches. We conduct experiments on the OU-ISIR Large Population gait dataset, the largest such database, and show that the proposed method outperforms state-of-the-art accuracy of generative and discriminative approaches under both identification and verification scenarios.