MicroCloud Hologram Inc. announced that it developed a point cloud denoising algorithm for the real-time 3D holographic reconstruction of single-photon LiDAR data. The algorithm is the result of the Company's independent research and development, which is conducive to further improving the Company's intellectual property protection system, maintaining its technological leadership, and enhancing its core competitiveness. Although 3D holographic LiDAR point cloud imaging continues to evolve rapidly, currently available computational imaging algorithms are often too slow, insufficiently detailed, or require extremely high arithmetic power, and even CNN-based (convolutional neural network) algorithms for estimating scene depth struggle to meet real-time requirements after training.

HOLO proposes a new algorithm structure that meets the requirements of speed, robustness, and scalability. The algorithm applies a point cloud denoising tool for computer graphics and can efficiently model the target surface as a 2D manifold embedded in 3D space. This algorithm can merge information about the observed model, such as Poisson noise, the presence of bad pixels, compressed sensing, etc.

This algorithm also uses stream modeling tools for computer graphics and can process tens of frames per second by selecting massively parallel noise reducers. HOLO's algorithm consists of three main steps: depth update, intensity update, and background update. Depth update: Gradient steps are taken for depth variables with point clouds denoised using the point set surface algorithm.

The update is operated in a coordinate system in 3D holographic space. Adaptation is performed on smooth continuous surfaces under the control of the kernel. In contrast to conventional depth image denoising, HOLO's point cloud denoising can handle an arbitrary number of surfaces per pixel, regardless of the format.

In addition, all 3D points are processed in parallel, resulting in short computing times. In addition, all 3D points are processed in parallel, significantly reducing computing time. Intensity update: Gradient steps are taken by targeting the coordinates of individual pixels in 3D holographic space to reduce noise.

In this way, only the correlation between points within the same surface needs to be considered. The nearest low-pass filter is used for each point. This step considers only local correlations and processes all points in parallel.

After the denoising step, points below a given intensity threshold, i.e., the minimum permissible reflectance, are removed.