Multivariate Fuzzy Density-Based Temporospatial Clustering of Applications With Noise (MF-DBTSCAN) Algorithm for Detection and Tracking of Multiple Targets by 77-GHz Millimeter Wave Radar Sensor

We propose multivariate fuzzy density-based temporospatial clustering of applications with noise (MF-DBTSCAN), a novel clustering algorithm combining multivariate, fuzzy, temporal, and spatial features. The primary objective of the proposed clustering algorithm is to establish well-defined clusters that address the challenges associated with proximity targets, crossing targets, and targets overlapping in range, Doppler, angle and their combination in multitarget environments. This objective is pursued through a thoughtful integration of spatial and temporal features alongside multivariate fuzziness to enhance the association of detections. In this proposed algorithm, range, Doppler and angle of arrival (AoA) in azimuth are used as three fuzzy variables with independent fuzzy membership functions to handle the probabilistic nature of the radar sensor measurements. Multivariate and fuzziness features of the algorithm are used to discern the clusters with discrete features, which eventually can be used to resolve the proximity targets and reduce track fragmentation and spurious tracks. The temporal feature of the algorithm includes dynamic clustering, a time-based minimal detection set during cluster formation, and purging of inactive tracks and false detections. This temporal feature significantly reduces cluster formation time and eliminates the false tracks. Well-formed clusters are fed to the Kalman filter (KF) for multitarget tracking. The proposed algorithm has been successfully validated with experimental data from a 77-GHz millimeter wave (mmWave) radar sensor mounted on a vehicle. Several experiments were carried out, encompassing a range of on-the-road scenarios involving radar mounted on a moving or stationary vehicle, moving and stationary objects, or stationary clutters. The scenario includes various situations where targets are in close proximity, including crossing ranges, having similar Doppler frequencies, close AoA, or a combination of these factors. Real-time data were collected with targets having different radar cross sections (RCSs), namely pedestrians, bicycles, motorcycles, and cars. The performance of the proposed MF-DBTSCAN is compared with state-of-the-art density-based spatial clustering of applications with noise (DBSCAN), fuzzy neighborhood-DBSCAN (FN-DBSCAN), and a density-based clustering method for multidimensional spatiotemporal data (MDST-DBSCAN).