| A traditional computer vision approach for autonomous driving as well as driver assistance is to detect all the objects (e.g., vehicles, pedestrians, buildings, trees, etc.) in the scene, and to infer that the area not containing any of these objects is the clear path. While such an approach may achieve a reasonable clear path detection performance, the overall system, which contains multiple object classifiers, is too complicated and too slow due to the need for multiple object detectors, probably one for each type of object.;In this thesis, we turn the problem around. We aim to detect the clear path directly in the scene. Our goal is to have only one clear path classifier instead of a combination of multiple object classifiers. We demonstrate our framework to detect the clear path directly using two different single camera-based vision algorithms: patch-based clear path detection and example based clear path detection. Moreover, by using other in-vehicle vision applications like lane mark detection and road boundary identification, we propose an application of clear path detection in identifying road shoulder on the highway. Since some weather conditions can severely degrade the appearance of the scene and lead to clear path detector failure, we propose a framework to extend the ability of clear path detection to tolerate such weather effects, in particular to deal with the effects of shadows and rain. Numerical experiments allowed us to evaluate the proposed framework in a quantitative way and show the results proving that proposed framework benefits the clear path detection and improves its detection performance under shadow and rain scenarios. |
