Multi-scale sparse convolution and point convolution adaptive fusion point cloud semantic segmentation method.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Semantic segmentation of LIDAR point clouds is crucial for autonomous driving systems.
• Existing methods face challenges with low segmentation accuracy and feature redundancy.
• Advanced feature extraction and fusion are needed for robust perception.

Purpose of the Study:

• To develop a novel approach for accurate semantic segmentation of LIDAR point clouds.
• To address feature redundancy and improve fusion of multi-scale features.
• To enhance the performance of autonomous driving perception systems.

Main Methods:

• Proposed an asymmetric importance of space locations (IoSL) sparse 3D convolution module to enhance sparse learning and feature extraction.
• Introduced a multi-scale feature fusion cross-gating module with cross self-attention for improved fusion accuracy.
• Utilized adaptive fusion of multi-scale sparse convolution and point convolution.

Main Results:

• The proposed approach significantly improves segmentation accuracy and robustness on SemanticKITTI and nuScenes datasets.
• Experimental comparisons show superior performance over state-of-the-art methods.
• Ablation studies validate the effectiveness of the individual modules.

Conclusions:

• The novel method effectively addresses limitations in current semantic segmentation techniques for LIDAR data.
• The adaptive fusion strategy enhances feature learning and fusion capabilities.
• The approach offers a promising solution for reliable autonomous driving perception.