Optimal LiDAR Data Resolution Analysis for Object Classification.

Area of Science:

• Geospatial technology
• Artificial intelligence
• Computer vision

Background:

• Accurate 3D object classification from LiDAR data requires efficient methods.
• Resolution of point clouds significantly impacts classification accuracy.

Purpose of the Study:

• To determine the necessary resolution for accurate 3D object classification using LiDAR data.
• To evaluate the performance of VoxNet with varying data resolutions.

Main Methods:

• Utilized the RedTail RTL-450 LiDAR System for data collection.
• Employed VoxNet, a convolutional neural network, for 3D data classification.
• Tested classification accuracy across different LiDAR data resolution levels.

Main Results:

• Training VoxNet with higher resolution LiDAR data achieved over 97% classification accuracy, even on sparse test sets (10% density).
• Classification accuracy dropped significantly when training with lower resolution data, particularly below 3% density.
• Higher resolution point clouds are valuable for both training CNNs and data acquisition.

Conclusions:

• High-resolution point clouds are essential for achieving high accuracy in 3D object classification with deep learning models.
• Findings inform optimal flight parameters (altitude, speed) for unmanned aerial vehicles (UAVs) in LiDAR data acquisition for classification tasks.