Camera-Assisted Radar Detection Clustering for Extended Target Tracking

Radar measurements from extended targets, such as vehicles or pedestrians, need to be clustered before they can be tracked by extended target tracking algorithms. However, clustering is challenging in autonomous driving scenarios, where multiple objects of different shapes and sizes remain in close proximity. Traditional clustering algorithms often produce wrong clusters due to merging and splitting, which degrade the pose estimation and tracking performance. In this work, we propose an approach to improve radar clustering using additional camera input. Our approach consists of two steps: first, we propose a range-based adaptive parameter selection method for radar-only clustering, which adjusts the clustering parameters according to the distance of the radar measurements. Second, we propose a camera-assisted clustering algorithm, which uses the camera images to refine the radar clusters by exploiting the visual information of the objects. We evaluate our approach by performing extensive simulations using the publicly available nuScenes dataset, which contains synchronized radar and camera data. Simulation results indicate that the proposed approach significantly improves the clustering performance in terms of accuracy and robustness.