Research On Pedestrian Re-identification Algorithm Based On Multi-feature Fusio

Person Re-identification(Re-ID)is a research task in the field of computer vision,which aims to recognize the same person across different camera views by using the collected pedestrian images.In practical applications,the appearance of pedestrians may change due to different times,locations,lighting conditions,as well as occlusion and posture variations,making Re-ID a challenging task.This paper conducted in-depth research on the core issues of Re-ID,including the exploration of relation feature mining,the lack of generality of algorithms,and the fusion of multiple feature potential relations.The specific work is as follows:(1)A Multi-Feature Fusion Network(MFFNet)was designed based on multi-level relation features,which introduced two new designs,the Feature Refinement Pooling Block(FRPB)and Feed-Forward Conduction Structure(FCS),to extract multi-granularity pedestrian features.In the FRPB,global and local features are fused,and multi-granularity features are extracted.The FCS utilizes different scales and resolutions of features for pedestrian matching.The results show that MFFNet achieved competitive performance on four datasets,demonstrating its effectiveness.(2)DAPNet employed a dual-stream branch design,including an attention pyramid branch and a mixed dilated convolution branch.In the attention pyramid branch,the model integrates multiple pyramid scales to obtain features based on the size and aspect ratio of the human body parts,and the Local Attention Module(LAM)provides accurate pixel-level attention to the feature descriptors obtained from the CNN model.The mixed dilated convolution branch adapts to semantic-spatial correlations by extracting features with variable receptive fields at different levels to counteract viewpoint differences.The results show that DAPNet achieved the best m AP/Rank-1 performance of 90.4%/94.3% and 85.7%/91.2% on the Market-1501 and Duke MTMC-re ID datasets,respectively,confirming its excellent performance.