Vision-based road boundary tracking system for unstructured roads

This paper addressed the problem to detect and track road boundaries of unstructured roads from images without road prior-knowledge. The unstructured roads pose the characteristic that their borders are not well defined but can be recognized either by color contrast like in rural walkways or by texture or line marks for example in urban bike paths. To have the ability to detect the road in the majority of cases, the system combines the detection of two modules, a color-based road detector, and a texture line detector. The color-based road detector works in the HSV color space and uses RGB normalization and white balance algorithms to reduce illuminations effects. Also, it uses a Manhattan distance measurement to segment the image into road pixels and a mean-shift algorithm to keep track of the road color characteristics. The texture line detector applies an inverse perspective mapping (IPM) to get rid of the perspective effects and detect parallel road borders based on a canny edge detector combined with a Hough transform. The robustness of the system is enhanced with the development of an Unscented Kalman Filter (UKF). The UKF estimates the best road borders from the measurements, also estimated borders in occlusion or in miss-detection and present a smooth road transition frame to frame. The system presents a recognition rate of 94% over 850 images within a variety of environments under various lighting conditions.