GRU-Based Kalman Filtering for 3D Multi-Object Tracking
Zikang Yuan,, Xiaoxiang Wang, Jiaxin Liu, Miaojie Feng, Zhaoxing Zhang and Xin YangB
AI summary
Problem
Traditional tracking-by-detection methods rely on fixed motion models and idealized Gaussian noise assumptions that cannot adapt to varying object dynamics, causing tracking errors and reduced precision.
Approach
The method replaces traditional Kalman filter equations with learnable GRUs that adaptively model state covariance and noise, trained via a hybrid-supervised strategy using pseudo-labels to handle sparse annotations.
Key results
- Automatically learns adaptive noise distributions and state covariances via GRUs
- Introduces pseudo-label-based hybrid-supervised training to overcome annotation scarcity
- Achieves 70.0% AMOTA and 265 identity switches on the nuScenes test set
- Releases open-source code for community adoption
Why it matters
Provides a robust, category-agnostic tracking solution that improves reliability for autonomous driving and robotics applications without manual model tuning.
Abstract
3D multi-object tracking is a crucial task for au- tonomous driving and robotics. Although tracking-by-detection- based approaches have demonstrated excellent performance in recent years, they ignore varying motion characteristics of different objects and utilize a single state space and estimator to model multiple categories. On the other hand, they model the noise of motion state as a ideal Gaussian distribution, which fails to capture the true complexity of the noise dynamics. These oversimplified modeling assumptions results in significant reductions of tracking precision. To address this limitation, we propose a learnable Gated Recurrent Unit (GRU)-based Kalman filtering, which is able to learn object motion char- acteristics through data-driven learning, thereby avoiding the necessity for manual motion and noise model design. To avoid abnormal supervision caused by the wrong association between annotations and trajectories, we adopt a hybrid-supervised learning strategy to accelerate the convergence speed and improve the robustness of the proposed method. Experimental results on two public datasets demonstrate that the proposed GRU-based Kalman filtering exhibits superior performance and significant potential compared to existing state-of-the-art approaches.