A Framework for Lane Prediction on Unstructured Roads

In this paper, we propose to address the issue of lane prediction on unstructured roads, i.e. roads where lane markings are not available. Lane prediction has received considerable attention in the last decade towards the development of ADAS (Advanced driver assistance systems) system. We consider lane prediction as a vision based problem and propose a learning based framework for lane prediction. We pre-process the data using adaptive thresholding to estimate ROI (Region of Interest) in an image. We develop a variant of Bayesian learning using the evidence based Cascaded Dempster Scafer Combination Rule to categorize the road and non-road sectors of the region of interest. We also propose to post-process the data with improved morphological operations to remove the outliers. Lane prediction finds applications in pothole detection, autonomous driving etc. We demonstrate the results on real datasets captured in different scenarios. We compare the results with different state-of-art techniques of lane prediction to validate the efficiency of proposed algorithm.