Map Outage Recovery: ICP Tolerance to Initialization Errors for Automotive Radar

This work presents a qualitative and quantitative analysis of the behaviour of the iterative closest point (ICP) registration algorithm when applied to the registration of scans from automotive radars to existing environmental maps. ICP is an algorithm commonly applied to point-cloud based navigation. While most often implemented with LiDAR data, radar sensors are increasingly prevalent as a sensors of use to localisation. ICP is a local registration algorithm, and relies on a reasonable initial pose estimate to avoid convergence to a local, incorrect minimum. However, position error can accumulate due to a number of reasons including drift in on-board motion sensors. Additionally, the first initialization of the ICP algorithm on start-up may contain errors. If ICP is to be used safely as a localisation algorithm, a good estimate of its resilience to initialization error is required. Road test data from radar and odometry sensors mounted to a vehicle navigating an underground parking garage is used for the analysis. A grid search method is applied to methodically test a range of positional and azimuthal offsets inflicted on the pose used to initialize the ICP algorithm. The ability of the ICP algorithm to then recover the true pose of the vehicle is examined across diverse scenarios is examined.