Graph-Based Spatial-Velocity Clustering and Tracking of Unlabeled Radar Point Clouds for Autonomous Vehicles with a Novel Temporal Precision Metric

Robust and accurate perception of the environment is crucial for autonomous driving systems. Radar sensors offer unique advantages over LiDAR and cameras, such as robustness to adverse weather, long detection range, and instantaneous velocity measurement via the Doppler effect, while being cost-effective. However, radar point clouds present challenges, including sparsity, noise, and ambiguous measurements. In this paper, we propose a graph-based framework for clustering and tracking unlabeled radar point clouds using velocity and spatial information through connected-component analysis. Our approach aggregates multi-sweep radar data into a unified global coordinate frame, enhancing spatial density and temporal continuity. We introduce a centroid-based temporal tracking algorithm that maintains consistent cluster identities across frames and multiple camera views. To objectively evaluate tracking performance, we develop the Multi-Attribute Temporal Consistency Precision (MATC-Precision) metric, which independently assesses temporal consistency using radar cross section and dynamic properties, features orthogonal to the clustering process. Experiments on the nuScenes dataset demonstrated that our method achieved over 80% tracking precision for clusters tracked across three consecutive frames. Qualitative results validated the robustness of our approach in complex urban environments, effectively handling occlusions, multi-camera views, and static and dynamic objects. This framework provides a scalable solution for reliable radar-based perception in autonomous driving.