A bayonet vehicle retrieval method based on camera perspective transformation

Abstract

The invention relates to a bayonet vehicle retrieval method based on camera perspective transformation. The present invention first performs feature point positioning on the bayonet vehicle image, mainly including the four corners of the front windshield, the four corners of the front engine cover, the headlights, the logo, the license plate, the front tire, and the like. Then use the prior knowledge of the vehicle feature points itself, such as the four corners of the front windshield are on the same plane, the four corners of the front hood are on the same plane, etc., to establish a reference plane for perspective transformation and a virtual camera model for perspective transformation. Finally, according to the established camera model, adjust the external parameters of the camera, and transform the image perspective into a fixed viewing angle, such as making the camera CCD plane parallel to the plane where the front windshield is located. The invention can reduce the vehicle angle of view error introduced by the difference of bayonets and greatly improve the retrieval performance.

Description

technical field [0001] The invention relates to a bayonet vehicle retrieval method based on camera perspective transformation. Background technique [0002] At present, most of the vehicle retrieval technologies with practical value are applied to road bayonets. The main reason is that the installation of bayonet cameras is standardized, thus ensuring that the captured vehicles are all head images, and the angle of view is within a certain range. . [0003] But even so, due to factors such as camera installation errors, lane differences, and vehicle driving directions, the vehicle images between different bayonets still have large differences in viewing angles, which leads to subsequent vehicle retrievals based on a certain viewing angle. Error, which is also the factor that has the greatest impact on performance in vehicle retrieval techniques. [0004] For this problem, there are few direct solutions in the existing technology, mostly in the vehicle feature extraction, t...

AI Technical Summary

Problems solved by technology

[0005] There are only a few patents, such as "201510081237.1", which deal with the viewing angle problem in a targeted manner, but this patent uses the simplest and violent multi-model method. In actual use, the same car is unlikely to have the opportunity to build a multi-view model. On the other hand, multiple models will also greatly increase the amount of data storage. For a small city with tens of millions of bayonet pictures per day, if multiple models are used to store feature data, the hardware cost will be greatly increased.

Images