SAFE-IMM: Robust and Lightweight Radar-Based Object Tracking on Mobile Platforms

Tracking maneuvering targets requires estimators that are both responsive and robust. Interacting Multiple Model (IMM) filters are a standard tracking approach, but fusing models via Gaussian mixtures can lag during maneuvers. Recent winnertakes-all (WTA) approaches react quickly but may produce discontinuities. We propose SAFE-IMM, a lightweight IMM variant for tracking on mobile and resource-limited platforms with a safe covariance-aware gate that permits WTA only when the implied jump from the mixture to the winner is provably bounded. In simulations and on nuScenes front-radar data, SAFE-IMM achieves high accuracy at real-time rates, reducing ID switches while maintaining competitive performance. The method is simple to integrate, numerically stable, and clutter-robust, offering a practical balance between responsiveness and smoothness.

💡 Research Summary

The paper introduces SAFE‑IMM, a lightweight yet robust variant of the Interacting Multiple Model (IMM) filter designed for maneuvering‑target tracking on mobile and resource‑constrained platforms, particularly using front‑facing automotive radar. Traditional IMM runs several motion models in parallel and fuses their posteriors via a Gaussian mixture. While this yields smooth estimates, the mixture update is computationally heavy and reacts slowly to abrupt maneuvers because the fused mean lags behind the true state. Recent winner‑takes‑all (WTA) schemes address responsiveness by selecting the most probable model outright, but they can cause discontinuities when the selected model’s state jumps far from the mixture.

SAFE‑IMM resolves this trade‑off by adding a covariance‑aware safety gate that mathematically bounds the drift that would occur if a WTA decision were taken. After a standard IMM prediction‑update cycle, the algorithm computes the posterior model probabilities w, identifies the winner w*, and maps all model states into the winner’s coordinate frame. Using the weighted Mahalanobis spread d² and the total tail probability t = 1 – w*, it defines a drift bound B = t·q·tr( P̄ )·d², where P̄ is a covariance average of the winner and the transformed rivals. If B ≤ ε_WTA (a user‑defined tolerance), the gate permits WTA and outputs the winner’s mean and covariance; otherwise the conventional Gaussian mixture (µ_mix, P_mix) is retained. The bound follows from a Cauchy–Schwarz inequality, guaranteeing that the Euclidean distance between the mixture mean and the winner’s mean never exceeds ε_WTA when WTA is allowed.

Beyond the gate, SAFE‑IMM incorporates two robustness enhancements: (1) Student‑t likelihoods (ν = 5) replace the standard Gaussian likelihood, down‑weighting large residuals and improving resilience to clutter and outliers; (2) Adaptive transition probability matrix (TPM) updates the Markovian model‑switching probabilities online using a blend of generalized likelihood ratio (GLR), entropy, and a winner‑streak bias. These adaptations accelerate model switching during maneuvers without sacrificing the stability of the filter.

Implementation-wise, the authors restrict the model set to two linear Kalman filters—Constant‑Velocity (CV) and Constant‑Acceleration (CA)—to keep computational load low. Multi‑object data association is performed with a Global Nearest Neighbor (GNN) tracker, avoiding the heavier JPDA or MHT approaches. The entire pipeline runs on a CPU‑only laptop (Intel Core Ultra 7 155H, 1.4 GHz) and achieves 17.9 FPS end‑to‑end (229.6 FPS for the IMM core alone).

Experimental evaluation consists of two parts. In synthetic simulations (Δt = 0.1 s, 30 s duration) three targets with distinct maneuver profiles are tested under two noise profiles: (i) high position noise, low velocity noise, and (ii) low position noise, high velocity noise. SAFE‑IMM consistently yields lower position RMSE and better Optimal Sub‑Pattern Assignment (OSPA) scores than a baseline IMM and a nonlinear EKF, while maintaining 100 % compliance with the safety gate (i.e., the drift never exceeds ε_WTA). The gate’s conservative behavior reduces ID switches dramatically, especially when velocity noise dominates and the filter prefers to stay in WTA mode.

The second part uses the nuScenes validation split, focusing on dynamic objects that are visible in the front‑camera view (≈13.8 k objects). SAFE‑IMM is combined with a sliding‑window radar detector and the GNN association. Compared with conventional IMM variants and with three monocular camera trackers (CenterTrack, Trac​e​S, PermaTrack), SAFE‑IMM achieves higher class‑agnostic AMOTA (7.24 % vs. 6.28 % for baseline IMM) and reduces ID switches by about 60 % (37 vs. 105). While recall is slightly lower due to the conservative gate suppressing marginal tracks, the overall tracking quality (MOTA, AMOTP) improves.

The authors conclude that SAFE‑IMM offers a principled solution to the responsiveness‑smoothness dilemma in maneuvering‑target tracking. By guaranteeing a bounded drift when switching to a single‑model output, employing heavy‑tailed likelihoods, and adapting the TPM online, the method delivers accuracy, robustness, and real‑time performance on modest hardware. Future work includes extending the framework to multi‑sensor fusion, incorporating non‑linear motion models, and automating the selection of gate parameters.