Real-time long-range road estimation in unknown environments

Real-time road estimation is important for dynamic path planning of autonomous driving, especially when construction sites or accidents occur along the road. Among the existing practices, main pain points are strongly dependent on lane markings and brightness. In addition, roads with a variable number of lanes are difficult to predict. To address these issues, we fuse the data sampled from camera, lidar, radar and predict the lanes 90m ahead of the ego vehicle in the top-view perspective by using a deep neural network. Our method is evaluated on real urban and highway scenes under good and bad weather conditions at different daytime, the results showed that our approach can handle roads with a variable number of lanes or roads with small T-intersections. Furthermore, the method can also deal with incomplete or missing lane markings at 166.66 fps.