A Method of Rough Localization of Multi-directional Vehicles in Static Images

Abstract

The invention discloses a multi-directional vehicle rough positioning method of a static image. Compared with the current vehicle positioning method based on a static image, the present invention has the advantage that the present invention can realize multi-directional rough vehicle positioning, and the vehicle's missing The detection rate is low; at the same time, the method in this paper transforms the multi-directional multi-vehicle positioning problem in complex scenes into the single-vehicle precise positioning problem in a small area, which provides a guarantee for further precise vehicle positioning.

Description

technical field [0001] The invention relates to the field of intelligent transportation, in particular to a multi-directional vehicle coarse positioning method for static images. Background technique [0002] Vehicle detection and positioning is the premise of multiple vehicle information positioning, identification, and analysis. It is a research hotspot in computer vision and intelligent transportation. At present, the video-based vehicle positioning method is relatively mature and can accurately locate the vehicle position. However, static images It is still difficult to guarantee the detection rate and accuracy in the current vehicle positioning method, especially for multi-directional vehicle positioning problems. [0003] At present, the vehicle positioning method based on static images is mainly realized by using classifiers, including methods based on pixel classifiers, such as Liu Huaiyu et al. (A color transformation model for static image vehicle detection [J]. Co...

AI Technical Summary

Problems solved by technology

[0002] Vehicle detection and positioning is the premise of multiple vehicle information positioning, identification, and analysis. It is a research hotspot in computer vision and intelligent transportation. At present, the video-based vehicle positioning method is relatively mature and can accurately locate the vehicle position. However, static images It is still difficult to guarantee the detection rate and accuracy of the vehicle positioning method, especially for the multi-directional vehicle positioning problem.

[0003] At present, the vehicle positioning method based on static images is mainly realized by using classifiers, including methods based on pixel classifiers, such as Liu Huaiyu et al. (A color transformation model for static image vehicle detection [J]. Computer System Application, 2010 ,19(9):191-194) first transform all pixel colors from the RGB three-dimensional color space to a two-dimensional color space, and then train the vehicle pixel classifier on the two-dimensional color space, but the classification of the classifier is accurate The rate is not high, and the position of the vehicle cannot be accurately positioned; tension, etc. (Application of color feature model in static vehicle detection [J]. Journal of Wuhan Engineering University, 2015,37(1):73-78) using Bayesian classifier The pixel separation of the road surface and the vehicle, and finally the separation of the vehicle target by the minimum cut / maximum flow algorithm, this method also has the problem of low positioning accuracy; and the classifier method based on the vehicle image, this method is suitable for the vehicle positioning problem in a single direction , can achieve a better vehicle detection rate, but the positioning accuracy is not enough, such as Li Xing et al. Using HOG+SVM for forward vehicle detection, the experiment proves that the detection rate can reach 96.52% under normal light; Chen Yangzhou (Robust vehicle detection algorithm based on Co-training method [J]. Computer Science, 2013, 39 (3): 394-401) proposed a vehicle robust detection algorithm based on the Co-training semi-supervised learning method. This method can only realize the detection of one-way vehicles, and the positioning accuracy is poor.

However, it is still difficult to solve the problem of multi-directional vehicle detection. If a classifier is used for detection, on the one hand, the multi-directional vehicle sample method is trained together, which is prone to overfitting and the positioning accuracy is not high. On the other hand, Because the size and proportion of the detection frame is uncertain, it needs to be detected by the multi-scale sliding window method, and the detection efficiency is very slow. Even if the classifier in multiple directions is trained, there will still be poor detection efficiency and screening problems.

[0004] In summary, for the multi-vehicle and multi-directional vehicle positioning problem in complex scenes, it is difficult to directly implement the classifier method. Therefore, the present invention proposes a static image multi-directional vehicle rough positioning method, which can find the location of the vehicle. roughly the area where it is located, so that the multi-vehicle and multi-directional vehicle positioning problem in a complex scene is transformed into a single vehicle positioning problem in a smaller area, and further vehicle positioning can be performed by other precise positioning methods

Images