Enhanced line-structured light 3D object detection based on a multimodal noise identification with feature fusion

Area of Science:

• Computer Vision
• 3D Sensing
• Geometric Processing

Background:

• 3D detection using line-structured light is crucial for spatial information capture.
• Point cloud data frequently contains noise from reflections and complex geometries, hindering accurate recognition.
• Existing methods struggle with small-scale noise removal and feature fusion.

Purpose of the Study:

• To develop an effective method for denoising 3D point cloud data.
• To enhance the performance of the FCAF3D framework for 3D detection.
• To improve the accuracy, speed, and surface quality of 3D reconstructions.

Main Methods:

• Proposed a multimodal noise identification network for small-scale noise removal.
• Integrated a CBAM-based multiscale feature fusion module into the FCAF3D framework.
• Optimized the loss function using CIOU (Complete Intersection over Union).

Main Results:

• Significantly enhanced point cloud denoising capabilities.
• Achieved improved 3D detection precision and increased inference speed.
• Produced smoother 3D reconstructed surfaces.

Conclusions:

• The proposed multimodal noise identification network effectively denoises 3D point clouds.
• The enhanced FCAF3D framework with CBAM and CIOU optimization improves detection accuracy and speed.
• The method offers superior performance for 3D structured light detection and reconstruction.