Abstract
This work studies the state estimation problem of a stochastic nonlinear system with unknown sensor measurement losses. If the estimator knows the sensor measurement losses of a linear Gaussian system, the minimum variance estimate is easily computed by the celebrated intermittent Kalman filter (IKF). However, this will no longer be the case when the measurement losses are unknown and/or the system is nonlinear or non-Gaussian. By exploiting the binary property of the measurement loss process and the IKF, we design three suboptimal filters for the state estimation, i.e., BKF-I, BKF-II and RBPF. The BKF-I is based on the MAP estimator of the measurement loss process and the BKF-II is derived by estimating the conditional loss probability. The RBPF is a particle filter based algorithm which marginalizes out the loss process to increase the efficiency of particles. All the proposed filters can be easily implemented in recursive forms. Finally, a linear system, a target tracking system and a quadrotor's path control problem are included to illustrate their effectiveness, and show the tradeoff between computational complexity and estimation accuracy of the proposed filters.
Introduction
In modern autonomous systems—ranging from GPS-guided navigation to remote target tracking—sensor data reliability is a critical bottleneck. Signal dropouts caused by urban "canyons," network congestion, or hardware failures often lead to intermittent measurement losses. While the Intermittent Kalman Filter (IKF) provides an optimal solution when loss timing is known, real-world systems often receive "garbage" data or pure noise without an explicit failure flag. This uncertainty transforms a standard linear Gaussian problem into a complex non-Gaussian, nonlinear estimation challenge where traditional tools like the Extended Kalman Filter (EKF) frequently diverge.
This paper addresses this gap by exploiting the discrete, binary nature of measurement losses (represented as $\gamma=1$ for success or $\gamma=0$ for loss). By targeting this binary switch specifically, the authors propose three recursive filters that balance estimation accuracy with computational feasibility.
Proposed Filtering Methods
To bridge the gap between simple but fragile linear filters and computationally prohibitive particle filters, three novel approaches are introduced:
- Bayesian Kalman Filter I (BKF-I): Utilizing a Maximum A Posteriori (MAP) approach, this filter makes a "hard decision" at each step regarding whether a measurement was lost based on data likelihood. It then applies the IKF update using this estimated state.
- Bayesian Kalman Filter II (BKF-II): This variant calculates the probability of measurement loss, performing a weighted "soft" update that accounts for uncertainty rather than committing to a binary decision.
- Rao-Blackwellised Particle Filter (RBPF): The most sophisticated of the three, the RBPF uses particles to represent the distribution of the discrete loss process while analytically marginalizing out continuous state variables. This hybrid approach is significantly more efficient than standard particle filters because particles are only required for the binary loss sequence.
Computational Complexity
A key advantage of the BKF variants is their efficiency. As shown in the comparison below, BKF-I and BKF-II maintain a computational complexity of $O(n^3)$, making them suitable for real-time embedded systems with limited processing power.
While the RBPF offers superior accuracy, its complexity scales with the number of particles $N$. To mitigate this, the authors also introduced a "Fast RBPF" implementation that removes duplicate particles inherent in discrete state spaces, significantly reducing the overhead for high-dimensional systems.
Experimental Results
The proposed filters were evaluated across three distinct benchmarks:
- Linear Stochastic Systems: Established a baseline against the "ideal" (but practically impossible) IKF.
- 2D Target Tracking: In nonlinear radar simulations, the RBPF consistently achieved the lowest Root-Mean-Square Error (RMSE). Notably, even when measurement loss probability reached 90%, the proposed filters remained stable while standard EKFs diverged.
- Quadrotor Path Control: In a high-fidelity Simulink model, the proposed filters allowed a drone to successfully navigate a triangular trajectory despite 20% packet loss in IMU and GPS data—a scenario where standard EKF-based controllers failed.
Conclusion
By specifically modeling the binary structure of sensor failures, BKF-I, BKF-II, and RBPF provide robust state estimation in environments with unknown measurement losses. While BKF variants offer EKF-like speed with improved robustness, the RBPF provides a high-accuracy alternative for more demanding applications. Future work will focus on online estimation of the loss probability parameter and establishing formal theoretical convergence proofs for these suboptimal filters.
Summary updated