Abstract
Precise environment recognition is a cornerstone of autonomous driving. In complex driving scenarios, vehicles must accurately track multiple dynamic obstacles, such as other cars and pedestrians, while overcoming sensor uncertainties. This paper proposes a multi-vehicle tracking algorithm using microwave radar. The system leverages Mixture Density Networks (MDN) to model the probabilistic distribution of radar measurements, specifically focusing on vehicle corners. To address the "curse of dimensionality" and manage a time-varying number of targets, the algorithm employs a Reversible Jump Markov Chain Monte Carlo (RJMCMC) sampling strategy within a particle filter framework. Simulation results demonstrate the robustness of the proposed approach.
Introduction
Tracking multiple objects (Multi-Target Tracking or MTT) is a fundamental challenge in intelligent vehicle systems. Radar sensors, while reliable in various weather conditions, produce noisy data that often clusters around specific vehicle geometries rather than the center of mass. Furthermore, the number of targets in a scene is dynamic, requiring a tracking framework that can adaptively adjust its state space dimension. Standard particle filters often struggle with these high-dimensional, variable-state problems.
Methodology
Radar Modeling via Mixture Density Networks (MDN)
Traditional tracking models often assume Gaussian noise centered on a point target. However, radar returns from vehicles are typically concentrated on the rear corners. To capture this multi-modal behavior, the researchers utilized Mixture Density Networks. The MDN outputs the parameters of a Gaussian mixture model (GMM) that represents the likelihood of radar measurements given a vehicle's pose. This allows the filter to prioritize returns from expected corner locations and effectively ignore environmental clutter.
RJMCMC Particle Filter
To manage the varying number of vehicles, the system integrates Reversible Jump MCMC into the particle filter's sampling step. Unlike fixed-dimension filters, RJMCMC allows the state vector to change size through "Birth" and "Death" moves.
As shown in the flowchart above, the process includes:
- Birth Moves: Proposing new vehicle states when unexplained radar detections appear.
- Death Moves: Removing existing tracks that are no longer supported by incoming data.
- Diffusion Moves: Updating the states of existing targets.
- Accept/Reject Step: A Metropolis-Hastings-like criterion ensures that jumps between different dimensions (adding or removing cars) are statistically sound, maintaining the stability of the posterior distribution.
Key Results
The proposed algorithm was validated through computer simulations involving multiple vehicles in various lane configurations and intersection scenarios.
The results highlighted several strengths:
- Robust Multi-Target Tracking: The system successfully maintained tracks on multiple vehicles simultaneously, even as they entered and exited the sensor's field of view.
- Clutter Rejection: By incorporating a Poisson-distributed noise model in simulations, the researchers demonstrated that the RJMCMC "Accept/Reject" mechanism effectively filters out false positives (ghost targets).
- Adaptive State Estimation: The "Birth" and "Death" mechanisms allowed the filter to automatically adjust the number of tracked objects in real-time without manual intervention or hard-coded limits.
Conclusion
By combining the expressive power of Mixture Density Networks for sensor modeling with the dimensional flexibility of RJMCMC, this research provides a robust framework for multi-vehicle tracking. The system effectively handles the non-Gaussian nature of radar returns and the dynamic complexity of real-world traffic. Future work will focus on real-vehicle implementation and the integration of additional sensors like LiDAR and cameras to enhance environmental perception.
Summary updated