A Car Window Segmentation Algorithm Based on Region Segmentation and Boundary Constraint
Abstract
To enable locate and segment the window of a car accurately in complex environment, a segmentation algorithm based on region segmentation and boundary constraint is proposed. At first, multi-scale and undirected graph based on region segmentation algorithm is applied to segment the vehicle graph into some small zones and sort according to set rules; then combine and merge the small zones in sequence, The merged region with the maximum boundary smoothness is served as the candidate window; finally judge for windows by combining with geometrical parameters. The experimental results show that the algorithm is accurate in segmentation, the result of the segmentation can maintain the shape and size of the window, therefore the fitness for purpose is strong and the application prospect is wide.
Keywords
Download Options
Introduction
With the rapid development of intelligent transportation system in China, the passenger detection plays an important role in traffic safety and traffic orderly operation. For example, the construction of HOV (High Occupancy Vehicle) dedicated lane is a feasible way to improve highway capacity. To guarantee the legal and reasonable use of HOV lanes, it becomes necessary to count the passengers in the vehicle. Again, the detection for wearing condition of the driver’s seat belt helps to promote the safety of driving, so it is particularly important to locate the driver’s position quickly. In addition, it allows as much as possible to observe physical characteristics and driving behavior of front row driver and passenger through the windows. Therefore, as the basic procedure of passenger detection, quick and precise locating car windows can significantly improve the speed and accuracy of passenger detection.
Car window location is a challenging issue, as the conditions of lighting, camera position and etc. in actual environment are complicate and variable, which makes it difficult to precisely segment vehicle windows from its body, meanwhile, due to the color depth of the body itself, part of the body color is difficult to distinguish with the windows, making it be more difficult for the segmentation of the window. Therefore, the focus of this paper is to segment and position windows accurately in complex natural environment.
There has been some researches for vehicle window location and extraction. Based on the characteristic that the edges of vehicle windows is closely linear, Hao et al.[1] proposed a vehicle window detection method based on linear characteristic, by carrying out differential operation to the adjacent frames of video image, this method detects upper and lower boundaries of the vehicle windows by combining horizontal linear filtering with Hough transform, obtains right and left boundaries by applying template matching, this method has a better effect of vehicle window detection for video sequence image, but the locating accuracy is not high when locating the vehicle windows of pictures with non-fixed time sequence and much bigger phase difference. Li et al.[2] proposed an window location algorithm based on Hough transform, which applies Hough transform to detect the horizontal straight line, and obtains two sides of the boundary to locate the window by further combining with the integral projection, this algorithm can enable quick locating of vehicle window with simple edges and clear image, however the location fails because it is difficult to find out the line segments when there are much more edges with the image in the vehicle window and the vehicle shooting angle is away from the positive direction. As the body color consistency is high, Wang et al.[3] made full use of the feature that the color difference of window region in the HSV color model is greater than that of the body to distinguish between the windows and the body, this algorithm has much better segmentation effect to vehicles with dark color. As the shape of the vehicle window is similar to that of a isosceles trapezoid, Hou et al.[4
Conclusion
This paper has analyzed the advantages and disadvantages of current vehicle window location algorithm, proposed a vehicle window segmentation algorithm based on region segmentation and boundary constraint. First the region segmentation shall be carried out for the car image according to the color feature, then the small regions will be combined and merged, finally the window region is obtained by combining the geometrical features of the vehicle window. The algorithm result not only locates the vehicle window accurately, but also keeps the shape and size of the window, which has laid a foundation for subsequent application of vehicle window image. The experiments show that the algorithm has better segmentation results for cars with different colors and different lighting conditions.