An Efficient Ensemble Deep Learning Approach for Semantic Point Cloud Segmentation Based on 3D Geometric Features and...

Area of Science:

• Computer Vision
• Geospatial Data Analysis
• Artificial Intelligence

Background:

• Mobile Light Detection and Ranging (LiDAR) sensor point clouds are vital for urban planning, autonomous vehicles, and infrastructure management.
• Accurate semantic segmentation of these point clouds is essential for extracting meaningful information and enabling advanced applications.

Purpose of the Study:

• To develop a robust and effective deep learning-based method for semantic segmentation of mobile LiDAR point clouds.
• To enhance the accuracy and efficiency of point cloud processing for 3D High Definition (HD) city mapping and related fields.

Main Methods:

• A novel deep learning approach, SegUNet3D, is proposed, combining U-Net and SegNet architectures.
• Mobile point clouds are transformed into range images using spherical projection for 2D representation.
• Local geometric feature vectors are computed for each point, and optimum parameters are determined through experimentation.

Main Results:

• SegUNet3D demonstrated superior performance compared to five other segmentation algorithms on both SemanticPOSS (urban) and RELLIS-3D (off-road) datasets.
• The method achieved significant improvements in mean Intersection over Union (mIoU), increasing it by up to 15.9% on SemanticPOSS and 5.4% on RELLIS-3D.

Conclusions:

• The proposed SegUNet3D method offers a highly effective solution for semantic segmentation of mobile LiDAR point clouds.
• This advancement has significant implications for improving the accuracy of 3D city maps and supporting autonomous vehicle navigation and urban planning.