Dynamic Die-Forging Scene Semantic Segmentation via Point Cloud-BEV Feature Fusion with Star Encoding.

Area of Science:

• Computer Vision
• Manufacturing Technology
• Artificial Intelligence

Background:

• Semantic segmentation of 3D point clouds is crucial for intelligent monitoring in hammer die-forging.
• Existing methods struggle with the deformation characteristics and pose variations common in forging environments.
• Current state-of-the-art (SOTA) approaches lack the specialized design for forging-specific challenges, impacting accuracy and real-time performance.

Purpose of the Study:

• To develop a novel semantic segmentation framework for complex die-forging scenes.
• To address limitations in capturing fine-grained deformation and multi-scale feature misalignment.
• To improve both segmentation accuracy and real-time efficiency for practical forging applications.

Main Methods:

• A framework fusing 3D point cloud and bird's-eye-view (BEV) representations.
• A Star-based encoding module in BEV encoding to capture workpiece deformation.
• Hierarchical feature-offset alignment and weighted adaptive fusion for cross-modal interaction and precision.

Main Results:

• The proposed method achieved a 1.1% higher mIoU than RPVNet when tested directly on real-world data after training on simulated data.
• Fine-tuning with minimal real data further improved mIoU by 5%, demonstrating significant performance gains.
• The framework effectively handles fine-grained deformation and multi-scale feature misalignment.

Conclusions:

• The novel fusion framework significantly advances semantic segmentation for hammer die-forging.
• The method demonstrates robust performance and adaptability across simulated and real-world forging scenarios.
• This approach offers a practical solution for intelligent process monitoring and quality control in forging industries.