Bearings-Only Tracking

• Bearings-only tracking is a nonlinear estimation technique that relies solely on bearing (directional) measurements to localize moving targets without access to range data.
• It employs advanced filtering methods—such as EKF, UKF, and particle filters—to manage severe nonlinearity and inherent observability challenges in passive sensing.
• Effective implementation requires strategic observer maneuvers like circumnavigation and trajectory optimization to enhance filter convergence and reduce estimation errors.

Bearings-only tracking (BOT) is a class of nonlinear filtering and estimation problems in which the observer is restricted to passive measurements of the bearing (azimuth, elevation, or directional cosines) of a moving target, without access to range or range-rate information. BOT is fundamental in naval, aerial, and robotic domains where stealth, sensor limitations, or operational constraints preclude active sensing or prohibit access to full spatial measurements. The core challenge arises from intrinsic lack of observability for static geometries, severe nonlinearity of the measurement map, and the necessity for observer maneuvers or cooperative sensing to induce informative estimation sequences.

1. Mathematical Formulation and System Models

A standard BOT configuration models the target state with position and velocity (planar or 3D), and the observer (ownship or agent) trajectory as deterministic or controlled. The canonical discrete-time target kinematic model in 2D is

$$w_{t+1} = F w_t + K \varepsilon_t,\quad w_t = [x_t,\,\dot{x}_t,\,y_t,\,\dot{y}_t]^\top$$

with $F = \begin{bmatrix}1&T\0&1\end{bmatrix}\otimes I_2$, $K = \begin{bmatrix}T^2/2\T\end{bmatrix}\otimes I_2$, $$\varepsilon_t \sim \mathcal{N}(0,\,\sigma_\varepsilon^2 I_2)$$. The observer state evolves via

$s_{t+1} = s_t + a_t,\quad a_t \in A(s_t).$

The only measurement made at each $t$ is the noisy bearing

$$\theta_t = \operatorname{atan2}(y_t - s_{y,t},\: x_t - s_{x,t}) + v_t,\quad v_t \sim \mathcal{N}(0,\,R)$$

or its extensions to elevation or azimuth/elevation pairs in 3D. No range information is available, leading to an unobservable system unless the observer executes a persistently exciting (and typically nonlinear) trajectory (Zhang et al., 2017, Li et al., 15 Aug 2025, Sciacchitano et al., 11 Feb 2026).

BOT extensions formalize the observer control problem as a Markov decision process (MDP), receding-horizon optimal control, or reinforcement learning (RL) over belief states represented by the filtering posterior (Zhang et al., 2017, Ristic et al., 4 May 2026). For multitarget and multisensor scenarios, bias modeling and fusion architectures are essential (Taghavi et al., 2016).

2. Nonlinear Filtering and Estimation Methods

BOT estimation is an archetype challenging nonlinear filtering scenario. Approaches include:

• Kalman-type Filters: The Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), and Cubature Kalman Filter (CKF) are routinely deployed with the nonlinear bearing measurement (Zhang et al., 2017, Singh et al., 20 Sep 2025, Ristic et al., 4 May 2026). The UKF/CKF are preferred due to improved performance in the strongly nonlinear update.
• Pseudo-Linear Estimation: Batch or recursive total least squares (RTLS) methods exploit the identity $\tan \theta_k = (y_k - y_{o,k})/(x_k - x_{o,k})$ and reduce BOT to an error-in-variables regression over pseudo-linearized data. Recursive RTLS, with careful modeling of noise in both measurement and observer state, substantially reduces bias versus pseudo-linear Kalman filters, especially when combined with circumnavigation to ensure observability (Li et al., 15 Aug 2025).
• Robust and Non-Gaussian Filtering: BOT with heavy-tailed or impulsive measurement noise requires robustification. The maximum correntropy criterion Kalman filter (MCC-KF) uses a kernel-based cost functional replacing the Gaussian likelihood, with bandwidth adaptation to minimize posterior covariance. This yields substantially reduced track-loss rates and RMSE under impulsive ocean noise (Singh et al., 20 Sep 2025).
• Particle and Evolutionary Optimization: For static or retrospective scenarios, parameter-based formulations (initial range, heading, speed) are optimized using evolutionary algorithms (GA, CMA-ES, PSO), with carefully constructed bearing-difference objective functions. CMA-ES converges orders of magnitude faster and more reliably than classical bit-string GAs in smooth cost landscapes (Kose, 2020, Xiao et al., 8 Jun 2026).
• Neural Network and Learning-Based Estimators: Recurrent neural networks, e.g., LSTM-based models, directly regress relative kinematics from sliding windows of bearing and ego-velocity, outperforming classical filters for rapidly maneuvering or nonholonomic targets, particularly in time-varying or unmodeled dynamic regimes (Torok et al., 28 Apr 2025).

3. Observer Control, Trajectory Optimization, and Stochastic Planning

The observer's maneuver is essential to resolving observability in BOT. Planning paradigms include:

• Circumnavigation and Persistent Excitation: Circular or oscillatory trajectories ("circumnavigation") induce persistent changes in relative bearing, ensuring the Fisher information matrix (FIM) is well-conditioned, leading to filter convergence and information-theoretic optimality (Li et al., 15 Aug 2025, Fu et al., 2024).
• FIM-Based Trajectory Optimization: Estimation-aware path planning optimizes Fisher information (D-optimality, spectrally-weighted log-determinant) to maximize informativeness and minimize the posterior Cramér–Rao lower bound across discrete time steps. For multi-platform systems, explicit intersection-angle (sine) objectives penalize degenerate (collinear) configurations, preventing aggregation and enhancing triangulation (Xiao et al., 8 Jun 2026, Sciacchitano et al., 11 Feb 2026).
• Dynamic Programming and Quantization: Fully discrete MDP approaches dynamically quantize the target state and observer action space, allowing Bellman recursions for trajectory optimization over finite horizons, including multi-objective tradeoffs such as detection efficacy and stealth (signal propagation) (Zhang et al., 2017).
• Reinforcement Learning: Deep RL formulates observer control as a belief-MDP using, for instance, DQN and CKF belief state, with reward balancing absolute error and filter consistency (Mahalanobis distance). RL-trained controllers match D-optimality benchmarks on mean accuracy but substantially reduce worst-case errors by regularizing against filter divergence (Ristic et al., 4 May 2026).
• Dual Objectives and Bi-Objective Control: Optimization-based BOT search, especially for Dubins or unicycle vehicles, frames the stepwise control law as a Pareto optimal tradeoff between D-optimal information gain and approach rate (range reduction), yielding closed-form solutions that outperform pursuit-only or estimation-only schemes (Li et al., 2019).

4. Geometric Observability and Enhancement Strategies

BOT is fundamentally limited by its sensitivity to geometry:

• FIM and Rank Conditions: The Fisher information matrix’s rank directly determines the observer's ability to estimate the full target state. Constant-velocity target motion with stationary observer or non-exciting trajectories induces a singular FIM, hence unobservable range/range-rate (Li et al., 15 Aug 2025, Sönmez et al., 2020, Sciacchitano et al., 11 Feb 2026).
• Visual Augmentation and "Bearing-Angle" Methods: In visual domain BOT, exploitation of bounding-box size to extract subtended angles ("bearing-angle" measurements) transforms classical rank-4 bearing-only estimation into full-rank (7D) observability (position, velocity, object length) with only minimal non-radial observer motion. This removes the necessity of lateral observer maneuvers and enables robust estimation even in "radial" pursuit (Ning et al., 2024).
• Multi-Agent and Networked Sensing: Fusion of bearing-only measurements from spatially separated observers (multi-UAV, sensor networks) dramatically improves geometric diversity and hence estimation performance. Distributed consensus, virtual fusion nodes, and bias estimation at the fusion node are central to network robustness and are provably convergent under minimal localizability (e.g., two non-collinear sensing agents) (Li et al., 23 Jun 2025, Huiming et al., 2023, Taghavi et al., 2016).
• Adaptive Planning and Learning: Gaussian process learning and planning construct agent trajectories maximizing the determinant of a "bearing-matrix" over a finite horizon, ensuring exponential estimation error contraction and adapting to unknown or time-varying target dynamics (Fu et al., 2024).

5. Alternative Representations, Robustness, and Higher-Order Methods

BOT performance and implementation can be improved by alternative coordinate systems and higher-order statistical reasoning:

• Log-Polar Coordinates (LPC) and Closed-Form Updates: Modified and log-polar representations make the measurements and state transitions more tractable, especially during instantaneous ownship maneuvers. Closed-form mean and covariance propagation (CFE-UKF) at maneuver points avoids sigma-point sampling and enables online computation of third and fourth central moments, providing real-time diagnostics of non-Gaussianity and initialization sensitivity (Xiourouppa et al., 2024). This enables adaptive re-initialization and triggers for switching filtering regimes.
• Maximum Correntropy and Non-Gaussian Filtering: For heavy-tailed, impulsive, or non-Gaussian measurement noise, kernel-based maximum correntropy filtering frameworks outperform Gaussian-based Kalman or particle filtering, when equipped with physics-based sensor array dynamic models and online kernel bandwidth adaptation (Singh et al., 20 Sep 2025).
• Global Optimization and Robust Cost Functions: Evolutionary optimizers (GA, PSO, CMA-ES) require well-chosen objective functions. The standard is the bearing-difference cost (direct measurement space residuals), which is robust and geometry-invariant, whereas the equidistant segment cost performs poorly under non-Gaussian distortion induced by polar-to-Cartesian transformations, as quantified by excess kurtosis in simulated distributions (Sönmez et al., 2020, Kose, 2020).

6. Applications and Implementation Considerations

BOT methodology is foundational in multiple domains:

• Underwater/Sonar Tracking: BOT is the backbone of passive submarine tracking, with emphasis on acoustic signal propagation, stealth, towed array dynamics, and multipath in uncertain marine environments (Zhang et al., 2017, Singh et al., 20 Sep 2025).
• Aerial and Ground Robotics: Coordinate-free BOT supports GPS-denied UAV operations, multi-robot formation, networked pointing, and cooperative circumnavigation, with direct applicability to autonomous surveillance, pursuit-evasion, and cooperative tracking (Sciacchitano et al., 11 Feb 2026, Li et al., 23 Jun 2025, Huiming et al., 2023).
• Vision-Based Pursuit: Bearing-angle augmentation and pseudo-linearization are especially valuable in image-based drone pursuit, enabling robust monocular tracking even without depth cues or stereo (Ning et al., 2024).
• Sensor Network Fusion and Bias Estimation: Distributed multi-sensor fusion and explicit, maximum likelihood bias estimation (via genetic algorithms, windowed batch estimation, and CRLB quantification) are essential to practical multi-agent deployments, securing position RMSE improvements up to 80% after bias correction (Taghavi et al., 2016).

Critical implementation aspects include observing geometry, excitation persistency (circumnavigation or planned trajectory), numerical conditioning (recursive filters or batch optimizers), hyperparameter tuning (forgetting factors, kernel bandwidths), estimator initialization (particularly range), and monitoring of filter divergence or non-Gaussianity. The tradeoff between estimation accuracy, control energy, observability, and platform constraints drives real-world system design.

---

References:

• (Zhang et al., 2017) Stochastic Control of Observer Trajectories in Bearings-only Tracking with Acoustic Signal Propagation Optimization
• (Li et al., 15 Aug 2025) A Recursive Total Least Squares Solution for Bearing-Only Target Motion Analysis and Circumnavigation
• (Torok et al., 28 Apr 2025) Bearing-Only Tracking and Circumnavigation of a Fast Time-Varied Velocity Target Utilising an LSTM
• (Singh et al., 20 Sep 2025) Bearing-only Tracking using Towed Sensor-Array with Non-Gaussian Measurement Noise Statistics
• (Ristic et al., 4 May 2026) Reinforcement Learning Trained Observer Control for Bearings-Only Tracking
• (Sciacchitano et al., 11 Feb 2026) Multi-UAV Trajectory Optimization for Bearing-Only Localization in GPS Denied Environments
• (Xiao et al., 8 Jun 2026) Trajectory Optimization in Single and Dual-UAV Bearing-Only Target Localization
• (Kose, 2020) Analysis of Genetic Algorithm on Bearings-Only Target Motion Analysis
• (Sönmez et al., 2020) Analysis of performance criteria for optimization based bearing only target tracking algorithms
• (Li et al., 2019) Optimization-based Control for Bearing-only Target Search with a Mobile Vehicle
• (Xiourouppa et al., 2024) An insightful approach to bearings-only tracking in log-polar coordinates
• (Ning et al., 2024) A Bearing-Angle Approach for Unknown Target Motion Analysis Based on Visual Measurements
• (Fu et al., 2024) Active Target Tracking Using Bearing-only Measurements With Gaussian Process Learning
• (Taghavi et al., 2016) Multisensor--Multitarget Bearing--Only Sensor Registration
• (Li et al., 23 Jun 2025) Networked pointing system: Bearing-only target localization and pointing control
• (Huiming et al., 2023) Bearing-based Simultaneous Localization and Affine Formation Tracking for Fixed-wing Unmanned Aerial Vehicles