Probably Unknown: Deep Inverse Sensor Modelling In Radar

TL;DR

This paper introduces a deep learning-based inverse sensor model for radar in autonomous vehicles, enabling accurate occupancy mapping and occlusion detection without human supervision, outperforming traditional filtering methods.

Contribution

A novel self-supervised deep neural network that converts raw radar scans into occupancy maps, incorporating uncertainty to identify occluded regions.

Findings

Outperforms standard CFAR filtering in occupancy segmentation

Successfully learns from partial lidar-based labels without manual annotation

Effectively quantifies uncertainty to detect occluded areas

Abstract

Radar presents a promising alternative to lidar and vision in autonomous vehicle applications, able to detect objects at long range under a variety of weather conditions. However, distinguishing between occupied and free space from raw radar power returns is challenging due to complex interactions between sensor noise and occlusion. To counter this we propose to learn an Inverse Sensor Model (ISM) converting a raw radar scan to a grid map of occupancy probabilities using a deep neural network. Our network is self-supervised using partial occupancy labels generated by lidar, allowing a robot to learn about world occupancy from past experience without human supervision. We evaluate our approach on five hours of data recorded in a dynamic urban environment. By accounting for the scene context of each grid cell our model is able to successfully segment the world into occupied and free space, outperforming standard CFAR filtering approaches. Additionally by incorporating heteroscedastic uncertainty into our model formulation, we are able to quantify the variance in the uncertainty throughout the sensor observation. Through this mechanism we are able to successfully identify regions of space that are likely to be occluded.