Singular Value Decomposition (SVD) Method for LiDAR and Camera Sensor Fusion and Pattern Matching Algorithm.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Sensor Fusion

Background:

• LiDAR and cameras are crucial for autonomous vehicles (AVs), offering complementary depth and visual data.
• Effective multi-sensor fusion is hindered by resolution, data format, and viewpoint differences.
• Existing calibration methods often require manual targets, which is impractical for dynamic AV environments.

Purpose of the Study:

• To develop a robust, targetless algorithm for aligning LiDAR and camera data in AVs.
• To address challenges in multi-sensor fusion caused by calibration drift.
• To improve the accuracy and reliability of sensor fusion for real-time perception.

Main Methods:

• A novel pattern matching algorithm utilizing Singular Value Decomposition (SVD) and Gradient Descent (GD).
• Alignment of geometric features (contours, convex hulls) between projected LiDAR point clouds and 2D image segments.
• Computation of an optimal transformation matrix for rotation, translation, and scale correction.

Main Results:

• Achieved up to 85% improvement in alignment accuracy on a vehicle-mounted platform.
• Reduced final projection error to less than 1 pixel.
• Demonstrated a practical solution for maintaining cross-sensor alignment despite calibration drift.

Conclusions:

• The proposed SVD-GD framework enables robust, targetless calibration for LiDAR-camera fusion.
• This approach provides reliable sensor fusion for autonomous driving applications susceptible to calibration drift.
• Enables real-time perception systems to operate robustly without the need for frequent recalibration.