Semantic Segmentation-Based Occupancy Grid Map Learning With Automotive Radar Raw Data

Precise road scene understanding is of great essence to autonomous driving. As a widely used method for road scene understanding, occupancy grid mapping is leveraged to detect obstacles and predict drivable road areas. Because of its robustness under harsh conditions, low cost, and large perceptual range, radar sensor is becoming increasingly important to achieve various critical perception tasks. However, for radar-based occupancy grid mapping, current inverse sensor model (ISM) relies on detection data and is most hand-crafted. In this work, we propose a novel data-driven ISM that employs the range-Doppler matrix as the input. With a systematic evaluation and comparison of our model with classic, hand-crafted ISM and the data-driven, detection-based Occupancy Net using RADIal dataset, we find that data-driven models are far superior to their hand-crafted counterpart. Furthermore, although both data-driven models are on par within near range ($<\! 50 \,\mathrm{m}$), our model outperforms Occupancy Net by a large margin in far range ($ [50 \,\mathrm{m},100 \,\mathrm{m}]$). Specifically, our model has about a 0.3 and 0.4 improvement in distant range for the intersection over union (IoU) and $\mathrm{F}_{1}$ score, respectively. In addition, leveraging adjacent occupancy grid map prediction, we propose a radar-based occupancy flow to precisely distinguish moving objects.