Abstract: Aiming at the research hotspot of cross-domain person re-identification (Person Re-ID) in video intelligent surveillance, the cross attention mutual learning for unsupervised cross-domain person re-identification is designed as a research experimental teaching project, and the influence of both the attention of important regional features and the mutual supervision of network pseudo tags on the cross-domain ability of the model is observed. A multi-scale cross combination attention mechanism is designed, and its enhancement of attention to important regional features in channel and space, as well as the reduction of attention to features with low discrimination is analyzed. On this basis, a multi-scale cross combination attention mutual learning experimental network is designed, and the training process is updated by hard label mutual learning to eliminate the impact of false labels on the performance of the model. Teaching practice shows that the experimental project helps to improve students’ practical ability and independent thinking ability, and provides a reference for the transformation of scientific research achievements into teaching experiments.