Vehicle re-identification method based on cross-domain migration enhancement representation

Abstract

The invention discloses a vehicle re-identification method based on cross-domain migration enhancement representation. The vehicle re-identification method comprises the steps: acquiring, optimizing and then iteratively updating a total target function based on a low-rank representation framework until a target function value converges; in the target domain, obtaining the mapping of the mapping ofeach sample relative to the user-defined vehicle attribute dictionary relative to the mapping of the discriminative feature dictionary shared by the source domain and the target domain relative to the unique discriminative feature dictionary of the target domain according to the picture feature vector of the new sample and each dictionary set obtained by iteration, and calling the mapping as a discriminative feature; linearly solving the attribute vector of the sample by utilizing mapping so that the new sample of the target domain can be expressed by the attribute vector and the discriminative feature vector zT of the new sample; and calculating cosine distances between attribute vectors of the to-be-matched vehicle sample and the candidate sample i in the target domain, calculating cosine distances between discriminative feature vectors of the to-be-matched vehicle sample and the candidate sample, summing the two cosine distances, and sorting {delta i} obtained by calculating all the candidate samples from large to small.

Description

technical field [0001] The invention relates to the technical field of intelligent transportation, in particular to a vehicle re-identification method based on cross-domain migration enhanced representation. Background technique [0002] With the development of cities and transportation systems, the installation of surveillance cameras is becoming more and more popular, and the problem of vehicle re-identification in multi-camera scenarios has attracted more and more attention from researchers. Vehicle re-identification technology can assist the police to lock suspect vehicles, which is of great significance for security and monitoring. Vehicle re-identification systems generally use vehicle images captured by monitoring systems with no overlapping fields of view. However, vehicle images obtained from different cameras often contain problems such as viewing angle changes, resolutions, illumination changes, and blurring. Come to great challenge. How to solve these problems ...

AI Technical Summary

Problems solved by technology

However, due to the differences between the vehicle data and the shooting equipment between the vehicle pictures in the source domain and the vehicle pictures in the target domain, the model trained in the source domain often does not perform well when applied to the target domain.

[0005] In addition, in practical applications, the target database may have problems such as small number of samples, lack of attribute annotation, and noise pollution.

Images