Abstract: A nested pooling tri-tuple convolution neural network is proposed for pedestrian re-identification, which is vulnerable to occlusion, illumination, viewing angle and other non-ideal conditions. It adds the average pooling and maximum pooling to extract global features in turn after the root mean square pooling, and automatically aligns the local features with the shortest path loss. Then the improved Log-logistic function is used instead of the traditional triple loss function to train the network, and get global features jointly optimized with local features. The recognition rates on the Market-1501, CUHK03, and VIPeR datasets are both more than 6% higher than those based on traditional methods. The experimental results show that nested pooling tri-tuple convolutional neural network proposed in this paper can effectively solve the problems of partial occlusion, low resolution and rotation change under non-ideal natural conditions, and has good generalization ability and scope of application.