Multi-branch Collaboration Based Person Re-identification

Abstract

As computer vision algorithms have advanced rapidly, the widespread deployment of video surveillance systems has enhanced traffic safety and propelled the growth of intelligent highway systems. However, the intricate nature of real-world scenarios, particularly the presence of occlusions, introduces noise that can cause the loss of critical feature information for identified individuals or objects. This poses significant challenges to current person re-identification algorithms. In response, this study introduces an innovative person re-identification approach that leverages a hybrid network architecture. The method performs feature extraction across four collaborative branches: a local branch, a global branch, a global contrast pooling branch, and an associative branch. This comprehensive approach yields a robust and diverse representation of person features, addressing the limitations posed by occlusion and noise in the environment. The neural network presented in this study is versatile and can be integrated with various backbone architectures. Empirical evaluations demonstrate that our proposed algorithm outperforms state-of-the-art techniques, while deep ablation studies substantiate the efficacy of the network’s design. This suggests that the architecture contributes significantly to the performance gains observed.