Deep Learning Powered Estimate of The Extrinsic Parameters on Unmanned Surface Vehicles

Yi Shen, Hao Liu, Chang Zhou, Wentao Wang, Zijun Gao, Qi Wang

Introduction

Autonomous navigation for Unmanned Surface Vehicles (USVs) relies heavily on the precise calibration of sensor extrinsic parameters—the spatial relationship between sensors like LiDAR and the vehicle's reference frame. While these parameters are typically static for ground-based robots, the marine environment introduces significant complexity. A USV's center of rotation, known as the metacenter, shifts dynamically in response to waves, buoyancy changes, and vessel motion. This paper proposes a deep learning approach to estimate these dynamic parameters in real-time, overcoming the computational bottlenecks of traditional physics-based simulations.

The Dynamic Metacenter Challenge

In naval architecture, the metacenter (M) is a critical point for determining a vessel's stability. As a ship tilts, the center of buoyancy (B) moves, and the intersection of the vertical line through the new center of buoyancy with the original centerline defines the metacenter.

Geometric representation of a ship's stability parameters, including the metacenter (M).

For high-precision sensor fusion, the extrinsic parameters must account for the shifting distance between the sensor and this moving metacenter. Traditional methods, such as Finite Element Method (FEM) analysis, can calculate these shifts with high accuracy but are computationally intensive, often requiring minutes to process a single frame. This latency makes them unsuitable for real-time autonomous navigation.

Methodology: Time-Sequence GRNN

To achieve real-time performance, the authors investigated neural network architectures as surrogate models for complex hydrodynamic physics. While standard Radial Basis Function (RBF) and General Regression Neural Networks (GRNN) can map Euler angles (roll, pitch, yaw) to 3D metacenter coordinates, they treat each time step as an isolated event.

The core innovation of this research is the Time-Sequence GRNN. Recognizing that the metacenter's movement is a continuous physical process, the authors introduced a feedback loop. The model takes two primary inputs: the current Euler angles and the predicted position of the metacenter from the previous time step.

The proposed Time-Sequence GRNN architecture incorporating temporal feedback.

This architecture allows the network to capture the temporal dynamics of the vessel's motion, leading to smoother and more accurate predictions compared to static regression models.

Experimental Setup and Data

Given the difficulty of obtaining ground-truth metacenter data in open water, the researchers utilized the Unity3D engine (powered by PhysX) to create a high-fidelity simulation environment. A virtual USV model was subjected to various sea states, and data—including Euler angles and metacenter positions—was sampled at 10 Hz. To ensure data quality, raw outputs were processed using 3D Gaussian and variance filters to mitigate simulation noise and outliers.

Results and Discussion

The performance of the Time-Sequence GRNN was benchmarked against three other architectures: a standard Fully Connected Network, an RBF Network, and a basic GRNN. Accuracy was measured using Mean Squared Error (MSE).

Network Structure Training MSE Testing MSE
Fully Connected Network 17.53 131.49
RBF Network 6.67 15.55
GRNN 6.41 14.33
Time-Sequence GRNN 3.10 6.65

The results demonstrate that incorporating temporal context significantly improves performance, halving the testing error compared to the standard GRNN. The testing MSE of 6.65 translates to a spatial error of approximately 2.58 cm. For a 9.5-meter USV, this precision is highly effective for maintaining sensor calibration during complex maneuvers.

Conclusion

This research presents a robust method for real-time estimation of dynamic extrinsic parameters on USVs. By leveraging a Time-Sequence GRNN, the system can approximate complex hydrodynamic shifts with high accuracy and minimal latency. Future work will focus on bridging the "sim-to-real" gap by validating the model with real-world sensor data and exploring more advanced temporal architectures like LSTMs or Transformers to further refine prediction stability.

Summary updated

BibTeX
@inproceedings{shen2024deep,
  author = {Yi Shen and Hao Liu and Chang Zhou and Wentao Wang and Zijun Gao and Qi Wang},
  title = {{Deep Learning Powered Estimate of The Extrinsic Parameters on Unmanned Surface Vehicles}},
  booktitle = {ACIRS},
  year = {2024},
  doi = {10.1109/acirs62330.2024.10684905},
}
Figure