GPU-Accelerated Point Registration:

Achieved a 10X performance improvement in the GPU-based ICP module for point cloud registration by utilizing KD-Tree for optimal nearest point pairing, learning point cloud statistics, removing outliers in the pairing process, and implementing other engineering optimizations.

3D Reconstruction Enhancement:

Increased point cloud generation speed by 20X through local plane fitting and hash checking, significantly expediting denoising and triangulation stages and reducing generation time from 170ms to 8ms. Enabled hole-filling functionality in the point cloud generation module, resulting in a final 3D point cloud with a three-dimensional resolution of 40 microns, capable of reconstructing gold metal surfaces.

SLAM, C++

Research Topic: Model-based Vision(estimate object pose in analysis-by-synthesis manner).

Advisor: Prof. Maxim Likhacehv.

Apart from CNN, I paid more attention to CUDA parallelization and reconstruction of the observed scene.

Conducted an efficiency depth image face detection algorithm which can work with low computational power based on MTCNN(Multi-task cascade CNN. Get over 97% recall and 99.8 precision on standard working conditions.

Detected the face Bounding Box by using the RGB image and the trained MTCNN model. Transformed the RGB face frame to depth image and saved as a face frame label for future 3D face detection algorithm testing.