RPC-Pillars: Radar Point Correction with Radar-PointPillars

Radar plays a big role in performing perception tasks for Autonomous Driving in adverse weather—such as during rainy, foggy, hazy, and snowy days. In the current era of rising demand for Autonomous Driving, the need for more accurate and versatile perception tasks are also on the rise. Perception task serves as a crucial stage in Autonomous Driving, as its outcome will be streamed down to other tasks such as trajectory prediction and obstacle avoidance. The topic of object detection has recently undergone a lot of investigation using camera and lidar. However, these sensors carry a limitation of performing poorly in adverse weather. Radar comes in handy in bad weather situations, but due to its sparsity, 3D radar object detection performs poorly. In order to tackle this challenge, we propose a novel radar point correction method to improve the radar point cloud density, followed by RPC-Pillars, a novel radar object detection algorithm. It establishes the new state of the art on nuScenes validation, achieving 2.8% higher mAP in the car class.