Dynamic Point Cloud Denoising via Manifold-to-Manifold Distance

Area of Science:

• Computer Vision
• 3D Data Processing
• Geometric Deep Learning

Background:

• 3D dynamic point clouds are crucial for applications like autonomous driving but are susceptible to noise.
• Existing denoising methods for static point clouds are insufficient for dynamic data due to complex spatio-temporal variations.

Purpose of the Study:

• To develop an effective denoising method for 3D dynamic point clouds.
• To address the challenges posed by irregular spatial and temporal sampling in dynamic point cloud data.

Main Methods:

• Representing dynamic point clouds using spatial-temporal graphs.
• Defining a manifold-to-manifold distance to measure temporal surface variations.
• Jointly optimizing the point cloud and graph representation with spatial smoothness and temporal consistency regularization.
• Developing an efficient algorithm for the optimization process.

Main Results:

• The proposed method significantly outperforms state-of-the-art static point cloud denoising techniques when applied frame-by-frame.
• Demonstrated superior performance on both Gaussian noise and simulated LiDAR noise scenarios.
• Effectively exploits temporal consistency for improved denoising accuracy.

Conclusions:

• The spatial-temporal graph-based approach offers a robust solution for dynamic point cloud denoising.
• This method enhances the quality of 3D dynamic point cloud data for various applications.
• The proposed technique provides a significant advancement over existing static point cloud denoising strategies.