Abstract: Vehicle re-identification is a technique for matching the same vehicle target under different conditions in a video surveillance system. In the case of vehicle re-identification, the images of the same vehicle in different cameras are different, and the single feature is difficult to describe the image stably, using a variety of features fusion characteristics of the vehicle, the method combines HSV and LBP features of vehicle images, and the fusion feature matrix singular value decomposition, feature extraction. In view of the slow convergence rate and low precision of traditional BP algorithm, the Levenberg-Marguardt adaptive adjustment algorithm is adopted to optimize the BP neural network. The experimental results show that the recognition rate of the method is 97.5%, and it is robust to illumination changes and angle changes.

Key words: feature fusion;vehicle re-identification;L-M adaptive adjustment algorithm;BP algorithm；singular value decomposition