Monocular Road Detection For Intelligent Unmanned Ground Vehicles

Good capability of environment perception and understanding is essential for Unmanned Ground Vehicles'autonomous navigation. In recent years, with the fast development of computer vision, vision-based road detection has become an important way to percept the environment. According to the different types of roads, the current road detection approaches can be broken down into two categories:structured road detection and unstructured road detection. In this paper, we proposed an approach for lane detection in structured road and an approach for road segmentation in unstructured road.Detection of the lane is the key for structured road detection. In this paper, a top view of the road is generated by using Inverse Perspective Mapping which is based on the camera calibration information. This will bring important benefit for the line detection since the lanes that appear to converge at the horizon in the image are now vertical and parallel. In order to reduce the noise, the IPM image is filtered by a two dimensional Gaussian kernel in both vertical direction and horizontal direction and followed by a thresholding. A simplified Hough Transform is used to count the lines numbers, and a RANSAC line fitting method to fit the lines. Based on the initial lines, a three degree Bezier spline is fitted to descript the lane. It shows good performance in the experiments on our own UGV in typical structured road and meets the real-time.Compared to structured road detection, unstructured road detection is a more difficult job. The traditional approaches for unstructured road detection usually focus on region segmentation based on color cues. These methods do not work well on the scenes with little different in color between the road and the offroad areas. In this paper, a texture-based method is proposed to detect the road boundaries. The vanishing point is estimated by a voting scheme for texture orientation which is calculated by Gabor filter, and the dominant two edged is detected by using LBP (Local Binary pattern), IN addition, a LBP+superpixels-based region segmentation approach is presented for the road where less texture can be found. The experiments demonstrate that it is effective for general road detection.