研究目的
To improve and adapt the standard unsupervised stereo camera calibration based on features matching to an RGB-NIR couple setup, dealing with feature matches across RGB and NIR images.
研究成果
The proposed unsupervised calibration framework based on SIFT flow and GA optimization effectively reduces global disparity, improves matching accuracy, and provides robust relative pose estimation for RGB-NIR pairs, making it suitable for computer vision applications requiring automatic calibration.
研究不足
The method assumes semi-calibrated stereo pairs with known camera matrices and no lens distortions. It may not handle large displacements or images with significant differences in resolution well. The simulation lacks depth information for parallax effects, and real-world applications like Kinect v2 require additional handling for distortions and resolution mismatches.
1：Experimental Design and Method Selection:
The framework uses SIFT flow for dense feature matching and incorporates a Genetic Algorithm (GA) optimization to minimize the global disparity field. It includes filtering stages to discard unreliable matches and a convex optimization for relative pose estimation.
2：Sample Selection and Data Sources:
The benchmark dataset from [7], synthetic scenes, and Kinect v2 data are used. The dataset includes RGB and NIR images, some registered and some transformed with projective transforms.
3：List of Experimental Equipment and Materials:
Kinect v2 sensor is used for real data capture, providing RGB, NIR, and depth map data. Cameras with filters for separating visible and NIR spectra are mentioned.
4：Experimental Procedures and Operational Workflow:
Features are extracted using SIFT descriptors, matched with SIFT flow, optimized with GA to reduce disparity, filtered to remove outliers, and used to estimate the fundamental matrix and relative pose via RANSAC and convex optimization.
5：Data Analysis Methods:
Performance is evaluated in terms of rotation and translation deviation from ground truth, using standard deviation metrics. Methods compared include SIFT, SURF, KAZE, SIFT flow, NRDC, and the proposed method.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
