Video-based Person Re-identification Algorithm Combining Spatial-temporal Self-attention Mechanism and Local Feature
Mingdong Yuan, Bin Wu, Hongying Zhang, Juntao Pu, Xue Li · 2023
Video-based person re-identification is susceptible to various interferences in video frames. This paper proposes a video-based person re-identification algorithm using the full-scale network (OSNet) that incorporates both spatio-temporal self-attention and local features to enhance the diversity of features. The proposed network has two branches, one for global feature representation and the other for local feature representation, which analyzes the features in the local branch using a blocking strategy and enhances the network's feature extraction capability by cascading the features of the two branches. Experimental results on the challenging video-based person re-identification dataset, MEVID, demonstrate that the proposed method outperforms most mainstream models and has a smaller model size than most existing methods.