LESA-Net: Semantic segmentation of multi-type road point clouds in complex agroforestry environment

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Point-cloud semantic segmentation is crucial for agricultural robots in agroforestry.
• Large-scale point-cloud data presents challenges in feature learning for segmentation.
• Accurate environmental comprehension is vital for robot navigation and operation.

Purpose of the Study:

• To develop an accurate point-cloud semantic segmentation network for large-scale agroforestry environments.
• To address the challenges of feature learning from massive point-cloud datasets.
• To improve the ability of agricultural robots to understand their surroundings.

Main Methods:

• Proposed a point-cloud semantic segmentation network framework utilizing double-distance self-attention.
• Introduced a local feature enhancement module extending receptive fields and enhancing feature generalizability.
• Developed a dual-distance attention pooling (DDAPS) block for aggregating discriminative local neighborhood features.

Main Results:

• The proposed network achieved accurate semantic segmentation of road-surface point clouds.
• Experimental results on SemanticKITTI and RELLIS-3D datasets demonstrated superior performance.
• The algorithm outperformed existing methods in large-scale agroforestry environments.

Conclusions:

• The double-distance self-attention framework effectively enhances point-cloud semantic segmentation.
• The proposed methods provide a robust solution for agricultural robots operating in complex environments.
• This work contributes to advancing the capabilities of autonomous systems in agriculture.