Neural Approach to Coordinate Transformation for LiDAR-Camera Data Fusion in Coastal Observation.

Area of Science:

• Robotics and Autonomous Systems
• Geospatial Science
• Sensor Fusion Technology

Background:

• Coastal observation requires accurate data on shore features.
• Unmanned Surface Vehicles (USVs) offer a platform for data collection.
• Sensor fusion, combining data from multiple sensors like cameras and LiDAR, enhances information quality.

Purpose of the Study:

• To develop and verify neural network algorithms for spatiotemporal alignment of camera and LiDAR data from a USV.
• To improve the accuracy and efficiency of coastal observation through sensor data fusion.
• To evaluate the performance of Multilayer Perceptron (MLP), Radial Basis Function (RBF), and General Regression Neural Network (GRNN) for coordinate transformation.

Main Methods:

• Utilized a point matching algorithm for coordinate transformation between camera and LiDAR data.
• Implemented and tested MLP, RBF, and GRNN neural networks for sensor alignment.
• Validated the proposed algorithms against real-world data recorded from a USV and compared with numerical methods.

Main Results:

• The proposed neural network approach demonstrated effectiveness as an alternative to traditional numerical methods for sensor alignment.
• Satisfactory accuracy was achieved for platform dynamics, with Root Mean Square Error (RMSE) often below 1 meter.
• While neural networks performed well on training data, generalization capability was limited, indicating challenges with matching point accuracy.

Conclusions:

• Neural network-based sensor alignment is a viable approach for enhancing coastal observation using USVs.
• Further research is needed to improve the generalization capabilities of the models and address limitations in matching point accuracy.
• Future work will incorporate vessel position and direction data to further refine the fusion process.