Fusion of Gait and Face for Human Identification at the Feature Level

Combining gait and face to identify humans can incorporate the advantages of both and improve the final recognition accuracy. Most of the previous work focuses on score-level fusion strategies. In this paper, we propose a multimodal fusion method to integrate information about gait and face at the feature level. Our approach separately concatenates the gait feature extracted by the GaitSet with the face feature extracted by ResNet50 (supervised by ArcFace loss), where the GaitSet and ResNet50 are trained in advance. The min-max normalization technique is utilized to transform the two biometric features to the common distribution space before concatenating, while a fully connected layer is used to further fuse the features after concatenating. To evaluate our approach, we built a multimodal gait-face database named CASIA-B-Gait-Face, which is based on the CASIA-B gait dataset. Extensive experiments show that our method achieves better performance than any individual biometric or other commonly used fusion methods.