Conventional DBSCAN algorithms exhibit limited effectiveness in complex traffic scenarios, particularly when vehicles are closely spaced or pedestrians are in close proximity to vehicles. Under such conditions, ambiguous object boundaries, poor cluster separability, and noise misclassification frequently occur. To address these challenges, this paper proposes an enhanced DBSCAN algorithm incorporating adaptive grid partitioning and velocity-based post-processing. First, logarithmic transformation is applied to preprocess the raw radar data. Then, a polar coordinate grid–based DBSCAN integrated with a dynamic elliptical neighborhood mechanism is employed to mitigate spatial distortion inherent in conventional methods. Subsequently, based on the statistical characteristics of Doppler velocity, outlier noise points are accurately removed using a velocity-consistency criterion. A spatial–velocity dual-similarity framework is further established to reassign misclassified noise points. Finally, a cluster-splitting strategy leveraging differences in speed and motion direction is introduced to resolve the issue of overlapping clusters in mixed-target environments. Experimental results demonstrate that, compared with the traditional DBSCAN algorithm, the proposed approach achieves substantial improvements in both clustering accuracy and noise suppression, providing an efficient and robust solution for millimeter-wave radar point cloud processing in complex scenes.