.png){kind=icon}
.png){kind=icon}
Return
Adaptive probabilistic multi-model framework for robust 3D multi-object tracking under uncertainty
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.126719.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
3D multiple object tracking is crucial for intelligent navigation systems and autonomous vehicles, enabling continuous localization and re-identification of surrounding objects. Current methods often rely on simple motion models like constant velocity and constant acceleration for trajectory prediction. However, these models struggle in uncertain environments where objects may perform abrupt maneuvers due to various factors such as collision avoidance, curiosity-driven exploration, or scene dynamics. Additionally, recent top-ranked trackers use category-specific motion models, slightly improving accuracy but increasing complexity and struggling with atypical object maneuvers. To address these challenges, we propose a dynamic probabilistic multi-model framework that leverages multiple parallel motion models for enhanced prediction accuracy, considering potential object maneuvers. Our framework introduces two novel motion models based on a learned probabilistic multi-category deviation angle. Combined with the constant velocity model, these yield three trajectory predictions, each assigned a probability of occurrence. We utilize a machine learning model based on logistic regression to estimate the probability of each prediction according to the object's momentum. Additionally, we introduce two probabilistic cost functions for updating trajectories during a two-stage data association process, prioritizing predictions with higher probabilities. A series of validation experiments were conducted on the KITTI and nuScenes tracking benchmarks. The results demonstrate that the proposed framework achieves robust tracking performance, outperforming many state-of-the-art trackers. Notably, it achieves a higher order tracking accuracy of 80.27% (cars) and 52.48% (pedestrians) on KITTI, and an average multi-object tracking accuracy of 75.5% on nuScenes. The code is publicly available at https://github.com/elhamesmaeeli/PMM-MOT.
Keywords:
Multi-object tracking
Motion modeling
Trajectory prediction
Logistic regression
Kalman filter
Computer vision