Aiming at the robustness problem in classical metric learning and the physical scene adaptability problem in current deep neural network of pedestrian re-identification

a method based on deep learning and metric learning is proposed. This method firsthy migrate and improve the structure and parameters of the pre-trained neural network model

and use data migration algorithm to augment every pedestrian sample of the CUHK03 pedestrian dataset. Then it uses the augmented pedestrian sampleset to fineiy tune the parameters of the deep neural network

and use obtain a neural network moder which adapt to pedestrian sample

and uses the deep feature extracted from a deep neural model

combine with the keep it simple and straight（ KISSME） metric learning algorithm to perform pedestrian re-identification. The test result shows that

Rank—1 reaches 55 % in standard database VIPe R and reaches 61 % in database Market—1501

which is more accurate than the original neural network model and related algorithms.