Roadside Detection Of Unstructured Roads

Lane edge detection is one of the most important tasks in image processing and scene analysis systems for mobile robots. In this paper, we present a review of techniques for the edge detection, and try to use them in our experiment. Most of them are developed for the gray image processing. In our experiment, the image captured by the ATRV-2 mobile robot is 24-bit. So the results are not satisfying. To get exact lane edges, we translate the (R, G, B) color space into (H, S, I) color space. And instead of using a fixed-triangle as a reference area, we provide a new method based on the deformable model used for reference. That is to say there is a road area in front of the robot, but the shape of the road area can change with the real scene. By making the hue and intensity values in the reference area into tow one-dimensional histograms, and comparing all pixels in the image to the hue and the intensity histograms, we can classify all pixels. It's really a fast algorithm. But when the model is not good enough, the result is not good either. So a more robust algorithm is needed. In this algorithm, (R, G, B) color space is translated into (Y, I, Q) color space, and the deformable model is also used. By using color attributes, mahalanobis distance can be calculated, and the road area can be extracted too. An improvement is that there are two constraint conditions. The one is considering the road area as the reference area only, and the other one is considering the road area and the roadside area both. Different constraint conditions can result in different lane edge image. So an algorithm used for choosing the better result is taken in this paper. And the better result chosen by the algorithm is considered as a standard edge result. It works well in our experiment, and the exact lane edges can be found in a short time. It is a quite useful algorithm for a mobile robot moving on a lane following a series of color image captured by itself.