An Improved Particle Filter for Target Tracking in Sensor Systems

Xue Wang, Sheng Wang, Junjie MaSensors 2007 · DOI

TL;DR

This paper introduces PF-RBF, an improved particle filter for target tracking that uses a Radial-Basis Function Network (RBFN) to dynamically adapt the target's motion model. This approach leads to more robust and accurate tracking, especially during target maneuvers, and offers a better trade-off between accuracy and computational cost compared to existing methods.

What Problem Does This Paper Address?

Sensor systems are crucial for surveillance, but they often struggle with accurately tracking targets, especially when those targets make sudden, unpredictable maneuvers. Traditional methods like the Extended Kalman Filter (EKF) can introduce significant errors in highly nonlinear scenarios. While generic Particle Filters (PF) can handle nonlinear and non-Gaussian problems, they suffer from particle degeneracy and accumulating estimation errors, leading to system divergence. Existing improvements like the Unscented Particle Filter (UPF) can be computationally intensive. The core challenge is to develop a target tracking system that is robust to maneuvers, accurate, and computationally efficient, particularly for systems requiring a large number of particles.

Methods

The paper proposes an improved particle filter called PF-RBF. The key innovation lies in its sampling step. Instead of relying on a static, predefined motion model, PF-RBF dynamically constructs and updates the process model using a Radial-Basis Function Network (RBFN).

Here's how it works:

  1. RBFN Training: The RBFN is trained using previous observations and current predictions to approximate the target's trajectory. This allows the model to adapt to the target's movement patterns.
  2. Dynamic Process Model: The trained RBFN effectively creates a dynamic process model that can better capture the target's motion, including maneuvers.
  3. Particle Sampling: Each particle's new state is sampled using this dynamically generated RBFN-based proposal distribution. This is a departure from traditional PFs that use a fixed proposal distribution.
  4. Standard PF Steps: After sampling, the algorithm proceeds with the standard particle filter steps: evaluating importance weights, resampling, and outputting the estimated state.

This RBFN-based sampling aims to provide a more accurate proposal distribution, reduce the accumulation of estimation errors, and maintain system convergence even during maneuvers.

Datasets

The paper does not use pre-existing datasets for training or evaluation. Instead, it generates synthetic data for simulation. The simulation involves a one-dimensional tracking scenario with a target exhibiting a maneuvering motion profile. The observation model is also non-stationary, switching between linear and nonlinear behaviors at a specific time step (k > 30). This synthetic data generation allows for controlled experiments to compare the proposed PF-RBF against other filters under specific conditions.

Key Insights and Ideas

  • Dynamic Motion Modeling is Key: The paper strongly suggests that a static motion model is a significant limitation in target tracking, especially with maneuvers. Dynamically adapting the model using techniques like RBFN is crucial for robustness.
  • RBFN as a Smart Sampler: The RBFN isn't just for prediction; its primary role here is to guide the particle sampling process. By approximating the target's trajectory, it provides a more informed proposal distribution, leading to better particle utilization.
  • Improved Robustness During Maneuvers: The PF-RBF's ability to adapt its motion model directly translates to significantly better performance when the target maneuvers, as evidenced by lower RMSEs during these periods.
  • Computational Efficiency Gains: While RBFN training adds some overhead, the paper argues that for a large number of particles, PF-RBF can become more computationally efficient than UPF due to its simpler update structure per particle.

Key Results

The PF-RBF algorithm demonstrated superior performance compared to the Unscented Kalman Filter (UKF), generic Particle Filter (PF), and Unscented Particle Filter (UPF) in their simulations.

  • Accuracy: PF-RBF achieved significantly lower Root Mean Square Errors (RMSEs) in both linear and nonlinear observation scenarios. Notably, during target maneuvers (k ≤ 30), PF-RBF's RMSE was substantially lower than the other algorithms.
  • Robustness: The algorithm proved robust for both linear and nonlinear observation models.
  • Execution Time: While the RBFN training step adds computation, PF-RBF's execution time scales more favorably with an increasing number of particles compared to UPF. The paper states that the rate of change of execution time for PF-RBF is about seven times less than UPF, making it a better choice for systems requiring many particles. For particle numbers exceeding 50, PF-RBF was faster than UPF.

Limitations and Future Work

  • Computational Complexity of RBFN: The paper acknowledges that RBFN training has a computational complexity of O(n³), where n is the number of training samples. However, they mitigate this by suggesting the use of only the nearest 'm' samples for training, which can significantly reduce complexity.
  • Critical Particle Number Thresholds: The paper notes that the "critical particle number thresholds" where PF-RBF becomes more efficient than UPF are problem-dependent and were found to be around 50 in their specific scenario. This implies that for different problems, these thresholds might vary.
  • Future Work: The paper implies that PF-RBF is well-suited for multi-target tracking scenarios that inherently require a large number of particles, suggesting this as a natural extension. Further exploration into optimizing the RBFN training and its integration with other advanced filtering techniques could also be future directions.

Summary updated

Abstract

Sensor systems are not always equipped with the ability to track targets. Sudden maneuvers of a target can have a great impact on the sensor system, which will increase the miss rate and rate of false target detection. The use of the generic particle filter (PF) algorithm is well known for target tracking, but it can not overcome the degeneracy of particles and cumulation of estimation errors. In this paper, we propose an improved PF algorithm called PF-RBF. This algorithm uses the radial-basis function network (RBFN) in the sampling step for dynamically constructing the process model from observations and updating the value of each particle. With the RBFN sampling step, PF-RBF can give an accurate proposal distribution and maintain the convergence of a sensor system. Simulation results verify that PF-RBF performs better than the Unscented Kalman Filter (UKF), PF and Unscented Particle Filter (UPF) in both robustness and accuracy whether the observation model used for the sensor system is linear or nonlinear. Moreover, the intrinsic property of PF-RBF determines that, when the particle number exceeds a certain amount, the execution time of PF-RBF is less than UPF. This makes PF-RBF a better candidate for the sensor systems which need many particles for target tracking.
BibTeX
@inproceedings{wang2007improved,
  author = {Xue Wang and Sheng Wang and Junjie Ma},
  title = {{An Improved Particle Filter for Target Tracking in Sensor Systems}},
  booktitle = {Sensors},
  year = {2007},
  doi = {10.3390/s7010144},
}
Figure