As is more robust compared to LIDAR, visual system, GPS and other proprioceptive sensors, radars operate well to detect objects at a long-range under different weather and lighting condition. The occupancy grid map is suitable for a high-level abstract representation of the environment, which is common in the robotics field. In this paper, we proposed a robust and precise radar-only algorithm of the occupancy grid map. Using the FMCW radar scan as an input, we first extract landmarks from noise measurements with minimal false positives and avoiding the detection of ghost objects. To estimate the radar’s relative motion, we associate with different landmarks corresponding to the same location at different times. With the pose information of radar and radar’s measurement, finally, we generate an occupancy grid map based on the inverse sensor model fit with radar’s characters. We experiment with our method in the Oxford Radar RobotCar Dataset, and the result shows map’s contours are sharp and distinct and our method can work well alone when other sensors drop out. Our grid map is also suitable for self-localization tasks that is precise enough for ego-motion estimation.

CICTP