Research On Vehicle Color Recognition In Natural Scene

Color is one of the most sensitive information of object for human vision. As for some unknown objects, human is able recognize their color firstly. In some special cases, human can discriminate the categories of the unknown objects. The color of the vehicle is one of the important information in urban traffic. The intelligent traffic system can obtain the registered information of the vehicle by plate recognition. However, for the reason of the small area of the plate region, the information of the plate is recognized correctly while the image is captured at high resolution. As the color of the vehicle is insensitive for the quality of the image, human is easy to recognize the color of the vehicle even for the low quality of the image. The recognition of the vehicle color is widely applied in the fields of criminal detection and law enforcement.Although the color of vehicle can be discriminated easily by human vision, the recognition base on compute vision may be difficult. Firstly, the illumination of the image influences the color easily. Different situations of the color offset even exist in the vehicles of the same color. The phenomenon of color offset causes the obstacle to color recognition. Moreover, the vehicle contains some structural information. Part of the vehicle region can be used for vehicle color recognition. Generally speaking, the vehicle color determined by the color of spray paint of vehicle, except for the windows, lamp, wheel, et al. For the image captured in the frontal view, the root and engine hood should be selected as the main region to determine the vehicle color. It is a difficult task to choose the interesting region from the images.In this thesis, we are aimed to solve the problem of vehicle color recognition under the natural environment. Specifically speaking, we focus on the representation of color information, the selection of the interesting region on vehicle and color feature learning.(1) The vehicle color recognition method under the nature environment needs to solve two problems. One is to overcome the influence of the different nature conditions for co lor. Another one is to choose the interesting region of vehicle for recognition, such as engineer hood or vehicle root. The traditional methods remove the color offset or the influence of light and convert the image under the canonical light condition in t he preprocess step. However, the preprocess step can tackle the situations under the nature scene hardly. In the paper, we propose a higher level feature which describes both the color of regions and the information of the distribution of the color features. The high level feature is more robust in the case of color offset. Moreover, the image is divided into small region. The importance of the region is determined by the weight of the feature of each small region. Due to the lack of the public dataset in the field of the vehicle color recognition, we publish a dataset contained 15601 images captured in the frontal view. The dataset is divided into 8 categories and contains different vehicle types, such as car, bus, truck, et al. The images under different situations of the nature scene, such as haze, strong light, are also included in the dataset.(2) Choosing the region to extract the color feature is the critical step in vehicle color recognition. The extracting of the discriminative region on the vehicle body is important in color recognition. In this paper, we propose a method based on the multi- instance learning algorithm to choose the region. Since the color distributions of the edge region and non-edge region are different, we divide the image into ed ge region and non-edge region base the edge constrained sample method. The discriminative region is able to select exactly by the combination of the two kinds of regions. As a result, the performance can be enhanced.(3) Color feature is fundamental for the vehicle color recognition. The feature is more helpful for recognition as its description ability is stronger. This paper proposes a feature learning method which learns feature from the low level to high level. Therefore, the feature is more discriminative.