Deep LiDAR localization using optical flow sensor-map correspondences

In this paper we propose a method for accurate localization of a multi-layer LiDAR sensor in a pre-recorded map, given a coarse initialization pose. The foundation of the algorithm is the usage of neural network optical flow predictions. We train a network to encode representations of the sensor measurement and the map, and then regress flow vectors at each spatial position in the sensor feature map. The flow regression network is straight-forward to train, and the resulting flow field can be used with standard techniques for computing sensor pose from sensor-to-map correspondences. Additionally, the network can regress flow at different spatial scales, which means that it is able to handle both position recovery and high accuracy localization. We demonstrate average localization accuracy of $\lt 0.04{\mathrm {m}}$ position and $\lt 0.1^{\circ}$ heading angle for a vehicle driving application with simulated LiDAR measurements, which is similar to point-to-point iterative closest point (ICP). The algorithm typically manages to recover position with prior error of more than 20m and is significantly more robust to scenes with non-salient or repetitive structure than the baselines used for comparison.