Abstract
Sensor registration is a fundamental prerequisite for effective multi-sensor data fusion. This paper addresses the challenge of calibrating 3D asynchronous multi-sensor systems characterized by unknown sensor biases and non-synchronized measurement intervals. We propose a computationally efficient estimation framework that leverages a nearly constant velocity (NCV) target motion model. By formulating the problem as a Weighted Nonlinear Least Squares (NLS) optimization and employing a two-layered approach—Block Coordinate Descent (BCD) combined with the Alternating Direction Method of Multipliers (ADMM)—the proposed method achieves near-theoretical accuracy with significantly lower computational overhead than existing state-of-the-art methods.
Introduction
In multi-sensor networks, such as radar systems, "sensor registration" ensures all sensors are spatially and temporally aligned. Uncorrected biases in range, angle, or sensor orientation (roll, pitch, and yaw) lead to significant errors in fused tracks, often manifesting as "ghost targets." Calibrating these systems in 3D is notoriously difficult due to two primary factors: the presence of nuisance variables (unknown true target positions) and high nonlinearity. Converting measurements from polar to Cartesian coordinates involves complex trigonometry that is further complicated by sensor orientation biases.
Methodology
The proposed method utilizes a Weighted Nonlinear Least Squares (NLS) formulation. A key innovation is the assumption of a nearly constant velocity target, which provides a physical constraint to link asynchronous measurements across different sensors without requiring prior knowledge of the target's true trajectory. This physical model allows the algorithm to effectively "cancel out" unknown target positions.
To solve the resulting non-convex optimization problem, the authors developed a two-layered approach:
- Block Coordinate Descent (BCD): This decomposes the high-dimensional problem into manageable sub-blocks, iteratively optimizing target velocities, range biases, and angular biases.
- ADMM for Non-convex Subproblems: Traditional methods use Semidefinite Relaxation (SDR) for orientation biases, which is computationally expensive ($O(M^{4.5})$). This paper adapts the Alternating Direction Method of Multipliers (ADMM) to solve these subproblems with a complexity of only $O(M^2)$, making it practical for large-scale systems.
Results and Discussion
The algorithm was evaluated using numerical simulations involving four sensors and a target trajectory across 20 random scenarios to ensure robustness against varying geometries and noise levels.
Key findings include:
- Accuracy: The estimator consistently reaches the Root Hybrid Cramer-Rao Lower Bound (RHCRLB), the theoretical limit of estimation performance.
- Efficiency: The ADMM-based BCD approach is approximately 40 times faster than SDR-based alternatives, reducing computation time from ~1300 seconds to ~33 seconds in the tested environment.
- Robustness: The method maintains high accuracy even under significant measurement noise and large initial bias offsets, outperforming traditional linearized filters that often suffer from model mismatch.
Conclusion
This work demonstrates that combining physical motion constraints with efficient optimization techniques like ADMM can solve the complex 3D sensor registration problem with high precision and speed. Future work will focus on extending the model to multi-target scenarios with clutter and utilizing quaternions to avoid mathematical singularities such as gimbal lock at extreme pitch angles.
Summary updated