Stereo Vision & Disparity

Overview

A solo computer-vision project walking the entire stereo pipeline in Python and OpenCV: from single and stereo camera calibration through rectification and disparity estimation, out to depth maps, 3D point clouds and optical flow on video.

Technical Highlights

• Camera calibration and undistortion. calibrateCamera on chessboard images recovers the intrinsic matrix and a five-coefficient distortion model, then undistort and remap correct the images. See main.py.
• Stereo calibration. Paired left and right captures give the extrinsics (rotation, translation, essential and fundamental matrices) at a mean reprojection error of 2.39 px. See Lab2/main.py.
• Rectification with epipolar lines. Image pairs are rectified and epipolar lines drawn on top to verify the alignment visually, also in Lab2/main.py.
• Three disparity methods compared. A custom 5x5 SAD block matcher alongside OpenCV’s StereoBM and StereoSGBM, all scored against ground truth with MAE, RMSE, bad-pixel percentage and SSIM, plus error heatmaps. See lab3.py.
• Depth and point clouds. Disparity is converted to depth from baseline and focal length, and reprojectImageTo3D exports colored PLY point clouds. See lab4.py.
• Motion analysis on video. Lucas-Kanade sparse tracking, Farneback dense flow, and morphological motion detection. See Lab5/main.py.

Learnings

Calibration turned out to be the most instructive part: intrinsics, distortion and extrinsics all have to be right before any disparity result downstream can be trusted.