MultiFG: a multimodal identity recognition method based on face and gait

Biometric recognition technologies leverage physiological characteristics (e.g., face, fingerprint, iris) and behavioral traits (e.g., gait, gestures) to perform individual identity authentication. However, single-modal biometric recognition heavily relies on high-quality feature extraction, where the presence of noise can significantly degrade recognition performance. By integrating multiple biometric modalities and exploiting their complementary information, recognition accuracy can be substantially enhanced. In this paper, we propose MultiFG, a multimodal identity recognition framework that combines face and gait information. Specifically, face images are extracted frame by frame using the MTCNN face detector, while gait skeleton sequences are generated via human keypoint detection with the MediaPipe pose estimator. Both modalities are processed through convolutional neural networks for feature learning. An attention mechanism is then employed to adaptively fuse the learned features, enabling robust multimodal identity recognition. Extensive experiments conducted on the MultiSubjects-Gait basketball action video dataset demonstrate that the proposed MultiFG method outperforms single-modality face or gait recognition approaches by effectively leveraging cross-modal feature complementarity, thereby achieving superior identity recognition performance.