Pedestrian reidentification based on multiscale convolution feature fusion

Kaiyang Liao, Gang Huang, Yuanlin Zheng, Guangfeng Lin, Congjun Cao

Abstract: The current pedestrian reidentification method based on convolutional neural networks still cannot solve the problems of pedestrian posture change, occlusion and background clutter. Many people use local feature learning or global feature learning alone to alleviate this problem, but they ignore their relevance. Aiming at the difference in emphasis between local features and global features, we propose a unified fusion algorithm, which inherits their advantages while discarding their shortcomings. While random erasure is used to enhance the robustness of the network model, the combined optimization function is used to optimize features of different scales, and the features processed at different scales are merged and spliced to obtain the final representation. Finally, multiple optimization reordering strategies are used to improve the performance of the algorithm. The proposed fusion algorithm was tested on three public pedestrian reidentification datasets, which proved the effectiveness of the method.