The video surveillance system collects video data in real time, which can serve as an effective third-party witness and provide favorable clues and information for case detection. However, due to the huge amount of data and the high retrieval workload, it brings inconvenience to the case collection. Aiming at this problem, this paper aims to retrieve the target person in the video evidence, and improves the existing cross-camera person weight recognition method to realize the rapid re-identification of the target person. Firstly, the target person image is segmented to obtain the block feature map; secondly, a local fusion module is introduced to fully retain local feature information and local correlation information; then a global fusion module is introduced to fully characterize the global image features while removing background noise; Finally, the cross-entropy loss and triplet loss function are integrated to accelerate the model convergence and effectively prevent over-fitting. The simulation experiment results show that compared with the existing methods of human weight recognition, the method in this paper has higher accuracy; the application software results show that the method in this paper can quickly locate the target person across cameras and meet the fast retrieval requirements of the target person in the video evidence.