Dempster Shafer Grid-based Hybrid Fusion of Virtual Lanes for Autonomous Driving

Lane estimation plays a central role in the context of self-driving vehicles, requiring clear road markings and a high quality infrastructure. To meet the safety requirements and manage difficult use cases (i.e. poor quality or absence of road- marking, traffic jams, severe weather conditions, etc.) without depending on cartography and lane markings, we propose a lidar-based Virtual Lanes (VL) generation system to provide a comfortable and safe ride. For this purpose, hybrid fusion based on Dempster-Shafer Theory (DST) coupled with a particle filter is introduced. Two VL strategies are merged: the first constructs virtual lanes based on independent road-borders detection, while the second uses moving objects trajectories. A novel lane similarity computation is also adopted in order to estimate an effective lane comparison. The performance is reported through extensive experiments with Valeo demo-car on highway and beltway roads. Experimental results demonstrate the accurate and robust performance of the proposed system.