Range-Doppler Registration

This paper presents a novel local self-localization approach which formulates the registration problem as a range-Doppler response problem. The proposed registration algorithm utilizes sensor observations in range-Doppler domain in combination with a pre-built high precision map and motion sensors for velocity and yaw rate measurement. In contrast to state-of-the-art approaches which perform the registration based on a point cloud, the introduced algorithm in this work shifts the registration from detection domain down to the sensor range-Doppler response. This procedure offers several positive characteristics such as independence of targets angle estimation, registration without any necessity of further processing of range-Doppler response for extracting detections and generating a point cloud. Furthermore, the transformation of the extracted detection from range-Doppler to Cartesian domain is not required.The registration problem is constructed as an optimization-based approach. The map information is converted to the range-Doppler domain and the optimization problem searches for the best overlap between sensor observation and map information. The registration result is subsequently used for correcting the ego-pose which is initially estimated based on motion sensors. The extensive simulation results demonstrate the feasibility and precise performance of the proposed algorithm.

I. Introduction

Autonomous driving systems demand a precise ego-motion estimation of the vehicle. E.g. accessing to the ego-pose information allows the controller to follow a planned route and calculates the possible deviations. It enables the controller to compensate these deviations by taking appropriate actions, such as accelerating, breaking or steering. Furthermore, tracking of pedestrians and objects such as driving vehicles in a digital map frame requires also a precise ego-pose information. In this work it will be assumed that an initial ego-pose (position and heading) estimation is available. This can be estimated by any positioning system such as mechanical motion sensors consisting of gyroscope(s) and wheel-based odometer(s) or a global navigation satellite system (GNSS). The systematic and non-systematic errors of the mechanical motion sensor systems lead to ego-motion estimation error which accumulate over time. Even a high precision differential global positioning system has several error sources which lead to a pose estimation inaccuracy. The registration process mitigates this inaccurate initial estimation by aligning the observations of a Doppler sensor to a pre-defined map. Digital maps such as OpenStreetMap (OSM) or sensor generated maps are being widely applied as an extra input for many ego-motion estimation algorithms. Concerning different digital map variants, the so called High definition (HD) maps contain, in contrast to the common digital maps, highly detailed and precise information of the environment. Certain features of these maps are used in the framework to be registered to sensor measurements. Fig. 1:

(a) A priori map containing poles (blue filled circles) overlaid with OpenStreetMap and the ego-vehicle. Dashed lines show some detected objects by the sensor. (b) The generated range-Doppler map for detected objects in (a).