Accurate feature point detection method exploiting the line structure of the projection pattern for 3D reconstruction

Area of Science:

• Computer Vision
• 3D Imaging
• Machine Learning

Background:

• 3D imaging with grid patterns offers real-time capabilities but struggles with accuracy due to scene inhomogeneity.
• Existing spatial coding methods face challenges in achieving high precision comparable to time multiplexing.

Purpose of the Study:

• To develop a more accurate feature point detection method for 3D imaging using grid patterns.
• To overcome the limitations of current methods in handling scene inhomogeneity and improving location accuracy.

Main Methods:

• A convolutional neural network (CNN) based approach was developed to detect feature points by analyzing the line structure of projected grid patterns.
• Two datasets were created to train the CNN model for extracting vertical and horizontal stripes from deformed grid patterns.
• A unique skeleton image was generated by fusing the predicted results from trained models for feature point detection.

Main Results:

• The proposed CNN-based method demonstrated higher location accuracy in feature point detection compared to previous techniques.
• The method effectively extracts and utilizes the line structure of deformed grid patterns for precise localization.
• Experimental results validated the improved performance in challenging imaging scenarios.

Conclusions:

• The developed CNN method offers a significant advancement in accurate feature point detection for 3D imaging applications.
• Exploiting the inherent line structure of grid patterns with deep learning provides a robust solution for enhancing 3D reconstruction accuracy.
• This approach paves the way for more reliable real-time 3D mapping in complex environments.