At present, the pedestrian re-identification method based on deep learning usually trained the classification network as the basic network, then used it to extract the deep features of the pedestrian images. Finally, calculated the similarities between features under the Euclidean distance metric and established a ranking table. Therefore, the feature representation ability of the classification network would affect the accuracy of re-identification, and the correlation between the feature representations would also cause errors in the similarity calculation. Aiming at these problems, this paper adopted a classification network with higher precision as the feature extraction network for pedestrian re-identification, and used singular value decomposition to reduce the correlation between weight vectors and optimized the deep learning process. In addition, it used the K-reciprocal encoding method to re-rank the pictures that need to be retrieved in the gallery that will further improve the accuracy of reidentification. Experimental results on the Market-1501 dataset show that this method can significantly improve the accuracy of pedestrian re-identification.

[2020-05-05] Printed Article /doc/2018.12.0918/v2