An Adaptive 3D Grid-Based Clustering Algorithm for Automotive High Resolution Radar Sensor

Novel automotive high resolution radar sensors can detect several thousands of reflection points from the surrounding environment, e.g., pedestrians, cyclists, vehicles and roadside infrastructure. For object classification and tracking, the detection points belonging to the same object shall be clustered into one group before further processing. This paper presents an adaptive clustering approach based on a range/angle/velocity-grid generated originally from the radar signal processing and angle estimation stage. In contrast to an x/y-approach, multiple reflection points will not be merged into one single grid cell at close ranges, but keep their individual information in different assigned grid cells. A time and storage efficient process with a clustering window according to grid indices is implemented to search for the points with similarity in all three dimensions. In order to eliminate the parameter dependency and the incorrect clustering due to uncertainties of real radar measurements, this approach is extended with a model-based clustering window depending on the tracked and estimated object contour. By validation with various measurement data, stable clustering results with almost perfect true positive rates are achieved independently of the prevailing parameters and object types.