Mutual Learning Person Search Based on Region Alignment

Person search is a combination of detection and re-identification tasks, which is widely used in the field of smart security. However, the current one-stage person search model without anchor has limited retrieval performance due to the lack of region alignment. In order to solve this problem, further research about the model structure and parameters has been done in our paper. In terms of model structure, we propose a mutual learning person search model based on region alignment, which adds a region propose branch on the basis of the person search model and improves the ability of region alignment for the feature extraction method. In terms of model parameters, we cluster the parameters of the network for the regions of interest. Experimental results show that the mutual learning person search model based on region alignment can effectively extract pedestrian features, and its performance exceeds that of the benchmark model while keeping the running speed.