Pedestrian Re-recognition Method Based on Pose Estimation and Regional Feature Fusion

Aiming at the local feature misalignment caused by the complexity of the pedestrian re-recognition scene and the difficulty of extracting the pedestrian features with invariance under the background clutter, we present a pedestrian re-recognition method based on the human body's gesture estimation and regional feature fusion. The pose estimation algorithm is introduced into the network to build aligned pedestrian features; The low-level global features are fused. Fine-grained pose features are constructed and aligned through two-branch structure in order that the model focuses more on key information about the target and the structural information of the human body. The correlation of global and local features is abstracted. The results of experiment on two mainstream datasets, Market-1501 and DukeMTMC-ReID, indicate that our method exceeds the level of the original baseline PCB algorithm. Compared with the baseline algorithm, the first hit accuracy Rank-1 is improved by 7.6% in the DukeMTMC-ReID dataset. The mean average precision mAP is improved by 12.8%.