A Novel Approach to Vehicle Pose Estimation Using Automotive Radar

Abstract

The accurate estimation of the pose, i.e. position and heading, of a vehicle while driving is of high importance in autonomous driving applications. Right now the main tool to estimate the location of a vehicle is its GPS sensor. However, GPS data is known to be of very low accuracy, especially in urban environments, and is thus not ideal for this application. In this report, the research into possible improvements to an existing vehicle pose estimation technique are presented, with the aim of making it applicable for automotive radar data. The existing technique is a scan-­matching technique known as the Normal Distributions Transform (NDT), which was originally designed for LIDAR measurements. By adapting the technique to accommodate radar measurements some of the drawbacks of LiDAR, such as the high cost and poor performance in certain weather conditions, can be overcome. Some of the main disadvantages of using radar as compared to LiDAR, e.g. lower resolutions, are addressed in the presented techniques. The lower resolution results in significant spreading of the target response, this is especially prominent in the azimuth domain for a standard 3 Tx × 4 Rx MIMO automotive radar system. By addressing the scan­matching problem in the polar domain, this spreading of the target response is better captured in the distribution used to perform the scan­-matching. Further, the implementation in polar coordinates allows for incorporation of the Doppler measurements, which contain knowledge about the angles of arrival of targets
and are generally measured at a much higher resolution than the angular measurements themselves. Moreover, the use of radar measurements introduces the availability of knowledge about the radar cross­-section of individual targets, this can be used to reduce the influence of false alarms. The incorporation of Doppler additionally allows the exploitation of the relation between the Doppler measurements and the angle of arrival to estimate a bias in the angle measurements, this can be used for sensor calibration while driving. The influence of the presented improvements to the NDT are examined through simulations and experiments. These results show significant reduction in the estimation errors. The sensor bias estimation technique also proves to provide extremely accurate estimates. Finally, a small extension is worked out to perform trajectory estimation using the individual poses by means of the Extended Kalman Filter.