Voxel-Based Neighborhood for Spatial Shape Pattern Classification of Lidar Point Clouds with Supervised Learning.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• 3D Lidar data classification is vital for autonomous vehicles and robots.
• Computing point neighborhoods in dense 3D scans is computationally expensive.
• Efficient spatial shape feature classification is needed for scene understanding.

Purpose of the Study:

• To propose a general framework for supervised learning classifiers using voxel-based neighborhood computation.
• To compare the effectiveness of different classifiers with this new framework.
• To evaluate the feasibility of voxel-based neighborhood for 3D lidar data classification.

Main Methods:

• A voxel-based neighborhood computation framework is proposed.
• Points within non-overlapping voxels are assigned the same class based on features in a support region.
• Five feature vector definitions using principal component analysis (PCA) were developed for scatter, tubular, and planar shapes.
• Neural Network (NN), Support Vector Machines (SVM), Gaussian Processes (GP), and Gaussian Mixture Models (GMM) were implemented and compared.

Main Results:

• The voxel-based approach significantly reduces computation time for neighborhood processing.
• The Neural Network (NN) classifier demonstrated superior performance.
• The framework proved feasible and effective across natural and urban environments using real-world 3D point clouds.
• Performance metrics and processing times confirmed the benefits of the NN classifier and the voxel-based method.

Conclusions:

• The proposed voxel-based framework offers an efficient solution for 3D lidar data classification.
• This method is highly beneficial for real-time applications in autonomous systems.
• The NN classifier combined with voxel-based neighborhood computation provides a robust and fast approach for spatial shape feature classification.