A Multi-Dimensional Attention Feature Fusion Method for Pedestrian Re-identification

Pedestrian re-identification aims to retrieve pedestrians across various cameras and scenes. However, the accuracy of re-identification is often affected by factors such as low-quality images of pedestrians and environmental conditions. Consequently, it is crucial for machine learning models to learn features from multiple dimensions. In response to these challenges, this paper proposes a Multi-Dimensional Attention Feature Fusion (MDAFF) method for pedestrian re-identification based on the NFormer approach. This method enables the model to learn and fuse pedestrian features from multiple dimensions, enriching the expressive power of the feature maps and improving the discrimination among pedestrians. By incorporating a PA module into the ResNeXt network for feature extraction, the model enhances its global perception and integrates pedestrian position information into the feature maps. This increases the model's sensitivity to pedestrian positions and reduces the impact of noise on re-identification accuracy. Furthermore, the method extracts channel and spatial correlations from the fused position feature maps and performs feature fusion, facilitating the fusion of multi-dimensional attention features. This alleviates the influence of varying scenarios and poses on re-identification, thereby enhancing the model's performance. Compared to the Res50+NFormer method, which directly models the relationships among different pedestrians after feature extraction, MDAFF integrates multi-dimensional features into the feature maps, improving the model's expressive power and capturing the relationships among different pedestrians more effectively. The proposed MDAFF method achieves a 1.3% increase in mAP and a 1.9% increase in Rank-1 on the Market1501 dataset, as well as a 1.7% increase in mAP and a 0.5% increase in Rank-1 on the DukeMTMC-reID dataset. Therefore, the MDAFF method effectively improves the accuracy of pedestrian re-identification.