Visual Lane Detection Algorithm Using the Perspective Transform

This manuscript develops a visual lane detection algorithm using the Hough transform to detect strong lines in a road image as candidates for lane boundaries. The search space in the Hough transform is reduced by searching for lane boundaries where they were detected in the previous video frame. The perspective transform is applied to determine the position and orientation of candidate lines, which are trusted as true boundaries if the detected lane width falls within a specified tolerance of the actual width. Results from a nearly 8minute long video of highway driving in rain indicate that lane boundaries are correctly identified in 95% of the images. Detection errors occur primarily during lane changes and poor lighting when entering underpasses. Including data from inertial measurements, location on digital maps, and steering direction would help to reduce or eliminate the instances of incorrectly detected lane location. INTRODUCTION Within the past decade the development of autonomous automobiles has grown significantly. Semi-autonomous vehicles are becoming more common in the commercial car industry. Audi, for example is reinventing their vehicles with many semi-autonomous features such as adaptive cruise control, side assist, and lane assist [1]. However, fully autonomous vehicles are still not available to the public. One factor of autonomous vehicles that requires refinement and improvement is the road detection system. The most common detection method is through a vision-based system due to the ability to collect data in a nonintrusive manner [2]. Typically a vision-based system uses a camera to capture footage and a sensor or laser to scan the preceding road area. Several methods are used to approach a vision-based road detection system. One concept analyzes road textures, and segments the road image based on road and non-road regions [3]. A system using visual memory stores guiding images to determine a drivable path for current road images [4]. A detection method for urban areas was proposed to identify curbs by using a Markov chain to detect and link curb points [5]. Issues involving these methods include the lack of precision and reliability when functioning on a variety of roads and the inability to identify lane boundaries on multi-lane roads. Lane detection can be accomplished using a variety of techniques. One system used color extraction to identify lane markings on roads. However size, shape, and motion was also considered in the detection process in order to differentiate lanes and cars of similar color [6]. A fusion-based method of laser scanners and video footage was developed to locate the drivable region and then detect any necessary lanes on roads [7]. A vision-based system incorporated vehicle localization on a digital map to detect and predict lanes [8]. Another approach develops modeled spatial context information shared by lanes, using an algorithm applied learns to address issues with shadows [9]. This manuscript introduces a lane-detection approach using the perspective transform, including an approximate distance from both the left and right lane boundaries. A detection algorithm is applied by searching for lines with the largest Hough transform value within the detection range where boundary lines are expected. The algorithm then determines whether the lines are to be trusted as the true lane boundaries based on their physical distances as computed from the perspective transform. In order to successfully navigate autonomously, only one line boundary need be trusted at any time as the position of other unknown boundary can be accurately approximated. PERSPECTIVE TRANSFORM Image generation in cameras follows the central imaging model in Figure 1 [10] where a camera with focal length f is physically placed at the origin of the XYZ axes. Using similar triangles, a point P in space is projected to position p on the xy image plane according to and X Y x f y f Z Z = = (1) A lane boundary line