The new trend of Multiple-Input Multiple-Output (MIMO) imaging radars in the automotive industry increases significantly the number of reflections in the point cloud, which demands low memory footprint solutions able to perform robust clustering with a low-latency. In this paper, we design, implement, and verify a tailored Density Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, dubbed Leaf-DBSCAN, whose heap memory is significantly reduced in comparison to the state of the art. The clustering strategy is modified to reduce the frequent distance queries between the target locations for cluster formation, and the clustering parameters adapt according to the spatial distribution of points. Leaf-DBSCAN is implemented in a RISCV-based processing element with constrained resources, and it is extensively validated using clustering standard datasets, and real data collected with a commercial 77-GHz Frequency Modulated Continuous Wave (FMCW) radar. The results display power and heap memory reduction in comparison to the traditional DBSCAN algorithm, while being able to process a high number of reflections in a small fraction of the frame time.