Lane Extraction and Tracking for Robot Navigation in Agricultural Applications

Abstract

In this paper, we propose a method for extracting and tracking man-made roads. It could be used for robot navigation in agricultural environments or hazardous related areas, as it is able to detect and track roads from images provided by an on-board color camera. Road extraction is achieved using color segmentation and principal areas detection; road tracking is achieved by active contours. First of all, our approach segments color images in small areas, which will be characterized later by color and texture attributes. These features are classified using the K-NN rule or the Support Vector Machines (SVM) method. A global scene model is obtained where the road extraction is used to initialize the active contour tracking process. Some of the road features are also used, in a focalized gradient zone in order to attract active contour to the road boundary. Besides, using the scene model information we can correct the transient errors generated in the tracking procedure. This algorithm has been evaluated on a great number of images, acquired either on secondary tarred roads or on earthen roads, mainly in countryside scenes. Results obtained on image sequences show the robustness of the proposed approach.