Multi-Scale Feature Fusion Network for Video-Based Person Re-Identification

In recent years, person re-identification technology has been greatly developed. Image-based person re-identification algorithms have achieved excellent performance on open source datasets. In contrast, the development of video-based person re-identification technology is relatively backward. At present, the main research work of video-based person re-identification algorithms is focused on the processing of temporal information in the picture sequence. Complex appearance features are not effective when performing temporal fusion, so the frame-level features used are almost based on global features. This paper proposes a video person re-identification model based on multi-scale feature fusion. The multi-scale feature fusion of the model is embodied in the design of the frame-level feature extraction module. This module extracts the frame-level features of different scales, and then catenates them together into vectors, which not only improves the feature discrimination degree, but also makes the catenated frame-level features carry out effective temporal fusion, and the test results on the Mars dataset have reached a competitive level. At the same time, a series of comparative experiments were carried out on the model parameters to achieve further optimization of performance.