Temporal Attention Framework Based on Occlusion Localization for Video Person Re-ID

Video-based person re-identification (Re-ID) refers to quickly locate the target pedestrians in multiple cross-device videos without overlapping, which plays an important role in the field of intelligent video surveillance. In the actual scene, occlusion will lead to the loss of local features of pedestrians, which will affect the discrimination of the target. Generally speaking, video has richer temporal and spatial information than images. It has significant advantages in addressing occlusion problems. In this paper, we propose a Temporal Attention Framework (TAF) based on occlusion localization. Different from the methods of key-frame screening or space-time completion, TAF focuses on extracting the attention of the same spatial region in temporal domains. Specifically, for one thing, we measure the correlation focusing of local features in multi-frame images to obtain the occluded part. For another, by compactly integrating the pairwise relationships of local features in the temporal domain with the features themselves, temporal attention is obtained. Experiments are carried out on the MARS and DuckMTMC-reID, and verify the proposed method has a good effect.