Unstructured Road Detection Based On Vision-Laser Data Fusion

For autonomous navigation of Unmanned Ground Vehicles (UGVs), we need process data acquired by multi-kind sensors to confirm measurable information of the road. Many practical methods are proposed to detect structured roads such as the highway. While unstructured roads have fuzzy boundaries and different kinds of surfaces which are sensitive to shifting of seasons and weathers. So long term stable road detection in them is difficult.To solve the problem, we design a vision system composed of two monocular cameras which are installed at different heights and angles. The one with a higher visual angle can cover remote area in front of the UGV. We can use the algorithm of Gabor filter to detect and track the vanishing point of the road. Then we can estimate the direction of the road. Considering the observation range of the onboard lasers, the other camera is set to cover9meters of road area in front of the UGV. For a picture captured by it, we use super pixel segmentation to divide it into200sub-images. Then a support vector machine algorithm is used to classify them into given classes. Moreover, we use morphological processing to improve the road detection accuracy. Because vehicles and pedestrians don’t have stable features for classification, the road detection algorithm based on monocular vision can’t assure safe driving of the UGV. On the basis of calibration of vision and laser data, we propose a kind of modeling approach that can present results of road detection in3-D laser data. We can adjust the road detection results with the help of geometrical characteristic in laser data. At the same time, we can describe the3-D road borders in the coordinate system of the UGV, so that the road detection results can be directly used in autonomous navigation.A self-build road database with more than10,000pictures is used to verify the effectiveness of the mentioned algorithm. The database covers pictures collected in spring, autumn and summer, which includes days of sunny, cloudy and rainy. The classification algorithm may fail in extreme weathers because features of those kinds of roads are quite different from samples used to train the classifier. So we proposed an online learning road detection algorithm to solve the problem. New samples collected from online acquired pictures are used to update the classifier in real time. Experiments on this strategy prove the modified algorithm can meet needs of autonomous navigation in unstructured environments.