A CNN-based Boundary Solver for Domain Decomposition Methods

• Developed a novel CNN-based Domain Decomposition Method (CNN-DDM) for efficient solution of interface degrees of freedom in large-scale fluid dynamics.
• Designed and trained a 3D CNN to replace the traditional Schur complement system, reducing computational overhead.
• Achieved <5 ms prediction time on GPU with <0.5% error in 3D lid-driven cavity flow simulations.

Computer Vision-based Obstacle Perception and Ranging

• Investigated stereo vision and deep learning approaches for obstacle perception in tunnel-like environments.
• Constructed and annotated a dataset of 850 images, training a YOLOv8 segmentation model for real-time obstacle detection.
• Implemented a C++ solution integrating Efficient Large-Scale Stereo Matching (ELAS) with a ZED 2 stereo camera, producing disparity maps (1344×376 @ 12 fps) for real-time ranging.

Disparity Map Generated with ELAS and ZED 2 Camera.

Wide-field Image Stitching for 3D Measurement

• Proposed a binocular stitching imaging model for multi-camera 3D measurement via binocular intersection solutions and implemented in a C++/Qt framework utilizing OpenCV.
• Validated system performance experimentally, achieving a 0.3153% measurement error rate in 3D measurements.