Crossing-Scene Pedestrian Identification Method Based on Twice FAS

摘要

In the field of crossing-scene pedestrian identification, the recognition accuracy is low due to the large local variation of the samples. A method based on twice Feature-Aggregation-Separation (FAS) is proposed in this paper. Firstly, a novel network structure aggregating the same types and separating different types of features twice respectively is proposed. Secondly, a method of cross-input neighborhood differences is applied to deal with the features produced by the first aggregation-separation, and the results are taken as the input of the second aggregation-separation. Finally, the features produced by twice FAS are chosen for splicing, and the results are used for Softmax classifier. Compared with MCPB-TC [8] method based on features aggregation-separation, the proposed scheme can provide directional aggregation-separation of positive samples and negative samples. Compared with AIDLA [4] based on cross-input neighborhood differences, it offers better ability of discriminating inter-class and aggregating intra-class. It also outperforms those methods by the tests of CUHK01 and VIPeR data set.