Lane-Aware Two-Stage Radar Multi-Object Tracking for Lane-Level Traffic Monitoring at Complex Intersections

Abstract

Reliable lane-level traffic monitoring at complex intersections is challenging with infrastructure radar alone: radar point clouds are sparse, cluttered, and provide only coarse angular resolution. We propose a lane-aware two-stage multi-object tracker that mitigates these limitations by combining generic tracking with lane geometry. First, a dual-clustering detector groups radar points into moving object hypotheses (using Doppler similarity) and supplements them with clusters centered on previously confirmed tracks, recovering temporarily stationary vehicles. Tracking then proceeds in two stages. To provide a stable starting point for lane-level analysis, the first stage runs an IMM-EKF in the global Cartesian frame. Building on these coarse estimates, the second stage distributes tracks among corresponding lane-specific trackers through probabilistic assignment, considering distance to the centerline, track covariance, and velocity heading. The track is converted into the Frenet frame and re-filtered, constraining the motion to the road axis and sharply reducing the lateral uncertainty. Because an independent tracker runs in every lane-specific Frenet frame, the system outputs not only refined track results but also real-time per-lane traffic counts. In CARLA simulations, our method significantly reduced velocity heading and lateral state RMSE compared to a standard Cartesian tracking baseline, while eliminating false track counts per lane. Real-world tests with a 24 GHz radar and RTK GNSS/INS ground truth show consistent improvements, substantially lowering both heading and lateral errors relative to the conventional Cartesian tracker. These validation results confirm that the proposed lane-aware two-stage tracker delivers reliable, lane-level traffic metrics even with noisy radar data. Comparisons with recent Frenet-frame trackers confirm our two-stage design achieves the most consistent accuracy at complex intersections where single-stage methods degrade.