Illuminant-invariant stereo matching using cost volume and confidence-based disparity refinement

Area of Science:

• Computer Vision
• 3D Imaging

Background:

• Illumination changes significantly degrade stereo-matching accuracy in 3D vision.
• Textureless regions pose challenges for accurate disparity estimation due to a lack of distinct features.

Purpose of the Study:

• To develop a robust stereo-matching method resilient to illumination variations and capable of refining disparities in textureless areas.
• To enhance the accuracy of 3D vision systems operating in radiometrically dynamic environments.

Main Methods:

• A novel stereo-matching approach combining an invariant image and Weber Local Descriptor (WLD) for cost volume computation.
• Utilizing a guided filter for cost aggregation and a confidence map for disparity refinement.
• Leveraging human visual characteristics for improved feature representation under varying light.

Main Results:

• The proposed method demonstrates improved disparity map accuracy in radiometrically dynamic environments.
• Significant reduction in disparity errors within textureless regions was achieved.
• The technique proves effective in overcoming challenges posed by large intensity differences between stereo images.

Conclusions:

• The developed illuminant-invariant cost volume and confidence-based refinement offer a robust solution for stereo matching.
• This method enhances the reliability of 3D vision systems, enabling applications in industrial robotics and autonomous vehicles.