Traffic Scene Semantic Segmentation Enhancement Based on Cylinder3D with Multi-Scale 3D Attention

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• 3D sensor technology advancements drive point cloud applications in autonomous driving, remote sensing, mapping, and manufacturing.
• Outdoor traffic scenes pose challenges for point cloud semantic segmentation due to disordered, uneven, and unstructured data.
• Traditional methods struggle with accuracy and stability in complex outdoor environments.

Purpose of the Study:

• To propose an improved point cloud semantic segmentation method addressing limitations of existing approaches.
• To enhance the model's ability to capture global and local features across multiple scales.
• To improve efficiency for processing large-scale point cloud data.

Main Methods:

• An improved point cloud semantic segmentation method based on Cylinder3D.
• Integration of PointMamba and MS3DAM modules for enhanced feature capture and multi-scale recognition.
• Incorporation of the KAT module into the encoder to boost perceptual capacity and robustness.

Main Results:

• The proposed method achieved a mean Intersection over Union (mIoU) of 64.98% on the SemanticKITTI dataset.
• Demonstrated a 2.81% improvement in mIoU compared to the baseline Cylinder3D method.
• Showcased superior segmentation accuracy and robustness in complex outdoor traffic scenes.

Conclusions:

• The proposed method significantly improves point cloud semantic segmentation accuracy over existing models.
• The integration of PointMamba, MS3DAM, and KAT modules enhances feature representation and processing efficiency.
• The approach offers a more robust and accurate solution for semantic segmentation in challenging real-world applications.