Semi-Supervised Segmentation Framework Based on Spot-Divergence Supervoxelization of Multi-Sensor Fusion Data for...

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Machine Learning

Background:

• Autonomous forest machines (AFMs) require robust perception systems for operation in complex, unstructured environments.
• Current multi-sensor fusion and segmentation methods often lack efficiency or accuracy for real-time AFM applications.

Purpose of the Study:

• To propose a novel semi-supervised segmentation framework for enhancing the intelligent capabilities of AFMs.
• To develop a method that effectively fuses multi-sensor data for improved object segmentation.

Main Methods:

• A semi-supervised segmentation framework utilizing spot-divergence supervoxelization of multi-sensor fusion data.
• Joint calibration of multi-sensor coordinates to create higher-dimensional fusion data.
• Gaussian density peak clustering for semi-supervised segmentation without manual parameter presets.

Main Results:

• The framework successfully balances supervoxel generation and semantic segmentation.
• Achieved high segmentation accuracy (F-score up to 95.6%) for various objects.
• Demonstrated efficient operation time, enhancing AFM performance.

Conclusions:

• The proposed framework offers a significant advancement in semi-supervised segmentation for AFMs.
• The method provides a robust and accurate solution for object detection in challenging forest environments.
• This technology has the potential to substantially improve the intelligence and operational efficiency of AFMs.