Semantic Segmentation Method for Sparse Point Clouds Based on Straight Flow Completion and Multi-Feature Fusion

Area of Science:

• 3D Computer Vision
• Point Cloud Processing
• Machine Learning

Background:

• Point cloud semantic segmentation is crucial but challenged by data sparsity and motion blur.
• Existing methods struggle with dynamic environments and moving objects, limiting practical use.

Purpose of the Study:

• To develop an effective semantic segmentation method for sparse point clouds, especially in dynamic environments.
• To improve the performance of point cloud completion and semantic segmentation models.

Main Methods:

• Integrated sparse point cloud completion with multi-feature fusion for enhanced segmentation.
• Developed efficient strategies for training point cloud completion models.
• Introduced a semantic segmentation model combining motion-enhanced instance and semantic features.

Main Results:

• The proposed end-to-end pipeline achieved accurate semantic segmentation of sparse point clouds in dynamic environments.
• Demonstrated superior performance in point cloud completion and semantic segmentation compared to classical methods.
• Validated on Lidar (SemanticKITTI) and radar (RADIal) datasets.

Conclusions:

• The novel method effectively addresses challenges of sparsity and motion blur in point cloud semantic segmentation.
• The integrated approach shows significant reliability and accuracy for real-world applications like autonomous driving.