Research on Pedestrian Re-identification Method Based on Visual Attention Mechanism

Abstract

Based on the method of attention mechanism, this paper researches on pedestrian re-identification. First, ResNet50 is used as the backbone network, and several model preprocessing methods are added as the baseline network. Then the channel attention mechanism module SENet is added to form the SE-ResNet50 network, and it can learn the importance of different dimensions of feature vectors, and focus attention on the corresponding dimensions. After the improvement, the model’s rank-1 on the Market-1501 data set increased by 1.5%, MAP increased by 2.0%, and the model’s rank-1 increased by 0.1% on the DukeMTMC-reID data set, MAP increased by 0.8%. In addition, this article also conducts a study on the importance of loss function, and the model obtains the best improvement effect when the ratio of ID loss to TriHard loss is 1 to 1.5. Rank-1 on Market-1501 increases by 0.2%, mAP increases by 0.2%, and on the DukeMTMC-reID data set, rank-1 increased by 1.1% and mAP increased by 2.1%. Finally, the actual video collection of campus scenes is carried out, and the trained model is applied to the pedestrian re-recognition in the actual scene. The results show that the model has an outstanding ability to recognize pedestrians in the context of a more complex environment.