Cross-camera pedestrian re-identification method based on Transform multi-head self-attention mechanism

The invention discloses a cross-camera pedestrian re-identification method based on a Transform multi-head self-attention mechanism, and relates to the technical field of target detection in computer vision, and the method comprises the steps: extracting continuous image frames in a monitoring video stream, detecting a pedestrian anchor frame in each frame through a target detector, and obtaining the confidence coefficient of each pedestrian anchor frame; an anchor frame area in the image frame is intercepted and sliced, and a slicing result is input into a hierarchical Transform encoder to obtain a feature vector of the pedestrian anchor frame area; sorting and screening the confidence value of each pedestrian anchor frame; and calculating the similarity between the feature vectors of the anchor frame areas of the two pedestrians, performing similarity comparison on the feature vectors of the anchor frame areas under different monitoring video streams, and clustering the feature vectors of the anchor frame a