Bayesian Approximation-Based Trajectory Prediction and Tracking with 4D Radar

TL;DR

The paper tackles robust 3D MOT for autonomous driving in adverse weather, where traditional LiDAR/camera systems struggle and Radar's resilience is enhanced by improved motion modeling. It introduces Bayes-4DRTrack, a 4D Radar MOT framework that combines a transformer-based motion predictor with Bayesian approximation in both detection and trajectory prediction, and employs a Doppler-informed two-stage data association. Key contributions include the first integration of Bayesian approximation for both detection and prediction in 4D Radar MOT, the Doppler-based two-stage association, and state-of-the-art AMOTA gains on the K-Radar dataset, notably an improvement of $5.7\%$. The approach yields greater robustness and accuracy in demanding, real-world driving conditions, with practical impact for safer autonomous navigation under adverse weather conditions.

Abstract

Accurate 3D multi-object tracking (MOT) is vital for autonomous vehicles, yet LiDAR and camera-based methods degrade in adverse weather. Meanwhile, Radar-based solutions remain robust but often suffer from limited vertical resolution and simplistic motion models. Existing Kalman filter-based approaches also rely on fixed noise covariance, hampering adaptability when objects make sudden maneuvers. We propose Bayes-4DRTrack, a 4D Radar-based MOT framework that adopts a transformer-based motion prediction network to capture nonlinear motion dynamics and employs Bayesian approximation in both detection and prediction steps. Moreover, our two-stage data association leverages Doppler measurements to better distinguish closely spaced targets. Evaluated on the K-Radar dataset (including adverse weather scenarios), Bayes-4DRTrack demonstrates a 5.7% gain in Average Multi-Object Tracking Accuracy (AMOTA) over methods with traditional motion models and fixed noise covariance. These results showcase enhanced robustness and accuracy in demanding, real-world conditions.

Figures (4)

• Figure 1: Structure of Bayes-4DRTrack. The proposed 3D MOT system consists of: (1) a Bayesian approximation–based 4D Radar detector, which produces object positions $\{d_t^i\}$; (2) a Bayesian approximation–based transformer motion prediction network, which outputs predicted tracks $\{\hat{trk}_t^j\}$; (3) a two-stage data association leveraging both Mahalanobis distance and Doppler velocity; and (4) a life-cycle management module for initializing/removing tracks.
• Figure 2: Bayesian approximation–based object detection results. Using MC Dropout, $N_D$ detection outputs are generated per frame. Bounding boxes are clustered via IoU; within each cluster, box parameters are averaged to obtain the final detection result, while standard deviation quantifies detection uncertainty. Notably, regions with high uncertainty often correspond to false alarms.
• Figure 3: Visualization of MC Dropout–based trajectory predictions. Blue dots represent the current object’s 4D Radar measurements, and the red box denotes the object's bounding box at time $t$. The black markers $p_t, p_{t-1}, p_{t-n}$ indicate past object states. Green points show the $N_P$ predicted positions for time $t+1$, whose mean and variance ($\mu_{\hat{p}_{t+1}}, \sigma^2_{\hat{p}_{t+1}}$) capture the estimated trajectory and uncertainty.
• Figure 4: Qualitative comparison of the CV model and the proposed prediction network. Shown are bounding boxes for matched tracks (in red), detection results, and ground truths over four consecutive frames (T, T+1, T+2, T+3) in a bird’s-eye view. Compared to the CV model, which accumulates errors when objects make abrupt changes (e.g., ID 2), the proposed prediction network accurately updates the track over time.