On-Board Video Based System for Robust Road Modeling

Abstract

In this paper, a novel road modeling strategy is proposed, defining an accurate and robust system that operates in realtime. The strategy aims to find a trade-off between computational requirements of real systems and accuracy and robustness of the results. The basis of the strategy is an adaptive road segmentation technique which ensures robust detections of lane markings and vehicles. A multiple lane model of the road is obtained by asserting hypotheses of lanes geometry based on perspective analysis and stochastic filtering. This multiple lane approach significantly improves vehicle location compared to other video-based works, as detected vehicles are accurately located within lanes. Tests show the adaptability, robustness and accuracy of the system in daylight situations with severe illumination changes, non-homogeneous color of the pavement of the road, lane markings occlusions, shadows, variable traffic conditions, etc., performing in real-time in all cases.