Feature detection for stereo-vision-based unmanned navigation

A fuzzy logic-based approach for the generation of feature maps with a commercial off-the-shelf (COTS) stereo-vision system is presented. This approach is applied for the guidance of autonomous vehicles in urban and outdoor environments. Useful features comprising obstacles and free-space regions are detected from disparity images generated by the COTS stereo-vision algorithm. Image pixels, disparity information and camera parameters enable the inverse mapping of each pixel and its associated properties such as colour composition onto an occupancy grid map that is used for navigation purposes. A measure of certainty in the location of obstacles within the perceivable environment is introduced by comparing the displacement of every grid cell from a dynamically estimated ground plane. In addition, discontinuities in the disparity images are identified as possible boundaries of features within the perceivable environment and mapped onto the same occupancy grid. Furthermore, a new fuzzy colour segmentation algorithm that quantifies the free-space within the perceived environment is presented. The advantages of this approach are computational speed and ease of implementation as well as obstacle detection robustness based on multiple features. Promising experimental results obtained on both live and recorded images demonstrate the effectiveness of this approach.