Research On Vehicle Shadow Detection Method Based On Color Attribute

With the development of technology and the progress of the times, computer vision technique, High-speed network communication technique and electronic technique are being developed gradually. Intelligent Transportation System plays a more and more important role in urban traffic planning and management. Meanwhile, the localization, recognition, speed measurement and tracking of moving vehicles in video sequence images are the main part of study. The first problem to solve in Intelligent Transportation System is to detect moving vehicles, and the detecting results would directly influence the later further processing. When detecting foreground, common object detection algorithms usually mistakenly detect shaded area as foreground. Whether removing the shadow or not rightly can directly affect the detection precision.After analyzing the forming principle of shadow and characteristics of color attributes, a shadow removing algorithm based on color attribute was proposed in this paper. Generally, light rays are parallel, so the closer the area nears the car, the darker the shadow is and the farther the area nears the car, the lighter the shadow is. Color attributes (color names) are learnt from our real life. Google image searching engine searches a cretin amount of image datasets for every color name, but datasets contain a larger number of error-positive samples. The proposed algorithm used probabilistic latent semantic analysis model (PLSA) to learn color attributes. Color attributes have the characteristic of mapping a normal image to a boundary clear, color clear image which without gradient pixels and can keep the completeness of the original image. According to this characteristic, the algorithm mapped darker shadow areas and vehicles to different color blocks. And after that, it carried binary processing on mapped image to meet the demand of removing the darker shadow areas.However, the processing of color attributes is concerned with the whole image, and it would mistakenly map neighboring area to foreground area. Thus, it would fail to locate the target and fail to remove shadow completely. In this paper, a background difference algorithm based on Gaussian Model was proposed. Firstly, in order to reduce the area size of foreground to be processed, it only detected moving vehicles as foreground. Secondly, it carried erode operation on the result of background subtraction to reduce the noise of the foreground edges. By using logical AND operator, it combined the result of background subtraction and the result of frame subtraction, which removed the impact of neighboring noises in the premise of ensuring the completeness of vehicles. To remove the influence of noise. Thirdly, in order to reduce isolated noises in edges or in inside regions, it used erode operation to remove them. And it used dilation operation to reduce detected holes inside. Finally, according to the connectivity characteristic of shadow areas, it filled area of a vehicle to obtain more accurate vehicle target.