Rural scene parsing and road boundary estimation by fusion of lidar pointcloud and EO images

This paper describes a perception system which fuses the output from a 3D lidar pointcloud classifier and an image scene parser to create a semantic 2D map of a rural, unstructured environment. This map is subsequently used to generate the dirt road boundary measurements, which are then fed into a particle filter to robustly track the dirt road boundaries in real time. A local road following reference path can be computed from the estimated road boundaries, which when combined with the semantic 2D map, can enable the robot to navigate in GPS challenged environments. The performance of the system is evaluated and verified through extensive experiments conducted out-field.