Isodisparity Profile Processing for Real-Time 3D Obstacle Identification J. Xu*, H. Wang**, J. I. Guzman*, T.C. Ng*, J. Shen*, C.W. Chan*** *Singapore Institute of Manufacturing Technology, 71 Nanyang Drive, Singapore 638075 **EEE, Nanyang Technological University, Singapore ***Defence Science & Technology Agency, Singapore ABSTRACT A new isodisparity profile based real-time obstacle detection algorithm is introduced for an autonomous vehicle. The input disparity map is preprocessed with a look-up-table to generate a number of isodisparity profiles which uniformly distributed over the terrain. For each isodisparity profile, row histogram analysis is performed to determine the reference line. Subsampling is then conducted to select only a limited number of profile points for obstacle and slope evaluation. By comparing the row coordinate of a profile point with the corresponding reference point, obstacles are detected with subpixel accuracy. The algorithm has several advantages over other methods: 1) 3-D obstacle detection can be achieved by 2-D isodisparity profile segmentation; 2) no conversion of disparity map to (x,y,z) image is needed; and 3) regardless the size of the input disparity map, only a predefined number of profiles and profile points are processed for obstacle map generation. The algorithm is 60 times faster than the conventional method [4] and is able to produce obstacle and slope maps in 4ms from a 320 x 240 disparity image (Pentium III, 850MHz PC). 1 INTRODUCTION Autonomous vehicles operating in off-road conditions must perceive the shape and surrounding terrain in order to navigate safely through obstacles. Perception must be both fast and accurate. Many researchers investigate 3-D obstacle detection using stereovision, primarily because a stereo system is able to provide dense 3- D data of the environments. The rapid advent of the computing power has generated a revolution in the way stereo systems are implemented and used. Examples of real-time stereo systems now abound in robotics. For instance, Digiclops from Point Grey Research [2] is able to generate 320 x 240 subpixel disparity map with a frame rate of 10 frames/second (PIII 850MHz). Another example is the SVM system from SRI [7], which also uses a conventional PC to deliver 320 x 240 16 level disparity map at 12 Hz. Equipped with dense 3-D data, the next task is to generate an obstacle map and a slope map for the navigation system to plan the path. There are a number of ways to detect obstacles from the disparity map. The fastest of obstacle detection algorithms, range differencing, simply subtract the range image of an actual scene from the expected range image of a horizontal plane. While rapid, this technique is not very robust, since mild slope will result in false indications of obstacles. The Antarctica autonomous vehicle Noman [4] from Carnegie Mellon’s Robotic Antarctic Meteoite Search program used 2 stereo pairs with auto iris, 3.6mm focal length lens. To reduce the cycle time, only a small number of rows in the image are examined by the stereo module. These rows correspond to distances of 4.5m to 8.5m in front of the robot. The disparity map is converted into (x,y,z) pixel coordinate to create a goodness (obstacle) map by using plane-fitting technique. For each cell in the goodness map, stereo fits a plane to the data in a region equal to the size of the robot (a 5x5 grid cell area) centering at the active cell. Smaller planes are also fit to each cell of this 5x5 submap. The goodness score of the center cell is then determined by the roll and pitch of the planes as well as the residual from the fitting planes. Thorpe, Hebert and Kanade [8] analyzed ERIM range data and constructed a surface property map represented in a Cartesian coordinate system viewed from top, which yielded surface type of each point and its geometric parameters for segmentation of scene map into traversable and obstacle regions. In our previous approach [3], we implemented a Hough transformation based obstacle detection algorithm. First, disparity map was transformed into height map in vehicle coordinate system. Stereo mismatched points in visible area were interpolated from the valid neighborhoods. Then, a global best-fit plane is calculated based on Hough transformation to provide general information about the surface