What the Cluster: An Unsupervised Framework to Segment Automotive Doppler Radar Pointclouds

The accurate identification of clusters from a radar point cloud is crucial for the safe operation of current Advanced Driver Assistance Systems (ADAS). Robust clustering of point clouds enables the correct identification and localization of other safety-critical traffic participants. Particularly, the case of undetected traffic participants presents an unsafe and dangerous situation. This paper introduces the Layered Doppler-Position Clustering (LDPC) framework for this task. In the first layer, static and dynamic detections are separated considering the translational and rotational movement of the vehicle. In the second layer, pre-clusters are developed based on one-dimensional hierarchical clustering using the Doppler velocity. In the final layer, each pre-cluster undergoes further clustering using DBSCAN and the position data. By separating the process into layers, LDPC effectively combines Doppler velocity and position data for clustering. Combining both variables with different units, such as meters and meters per second, in a distance function, violates the semantic integrity of the data. Furthermore, the experimental results shown in this paper prove the applicability of the proposed LDPC framework. In various driving situations, the accuracy of the clustering results can be improved compared to existing methods. Additionally, we consider the safety-critical limits concerning specific radar sensor errors. Across all validation data, we achieved higher accuracy compared to existing systems in our experimental results.