Pedestrian Re-Identification Based on Fine-Grained Feature Learning and Fusion

Video-based pedestrian re-identification (Re-ID) is used to re-identify the same person across different camera views. One of the key problems is to learn an effective representation for the pedestrian from video. However, it is difficult to learn an effective representation from one single modality of a feature due to complicated issues with video, such as background, occlusion, and blurred scenes. Therefore, there are some studies on fusing multimodal features for video-based pedestrian Re-ID. However, most of these works fuse features at the global level, which is not effective in reflecting fine-grained and complementary information. Therefore, the improvement in performance is limited. To obtain a more effective representation, we propose to learn fine-grained features from different modalities of the video, and then they are aligned and fused at the fine-grained level to capture rich semantic information. As a result, a multimodal token-learning and alignment model (MTLA) is proposed to re-identify pedestrians across camera videos. An MTLA consists of three modules, i.e., a multimodal feature encoder, token-based cross-modal alignment, and correlation-aware fusion. Firstly, the multimodal feature encoder is used to extract the multimodal features from the visual appearance and gait information views, and then fine-grained tokens are learned and denoised from these features. Then, the token-based cross-modal alignment module is used to align the multimodal features at the token level to capture fine-grained semantic information. Finally, the correlation-aware fusion module is used to fuse the multimodal token features by learning the inter- and intra-modal correlation, in which the features refine each other and a unified representation is obtained for pedestrian Re-ID. To evaluate the performance of fine-grained features alignment and fusion, we conduct extensive experiments on three benchmark datasets. Compared with the state-of-art approaches, all the evaluation metrices of mAP and Rank-K are improved by more than 0.4 percentage points.