Stereo Matching and Optical Flow Field Computation

Stereo matching and optical flow field computation represent classical problems in computer vision, widely applied in domains such as 3D reconstruction, motion analysis, and robotic navigation.

The core objective of stereo matching is to identify corresponding pixels of the same scene from two different viewpoints, thereby calculating a disparity map to recover depth information. In MATLAB implementations, this is typically achieved through block matching algorithms (e.g., Sum of Absolute Differences - SAD, Sum of Squared Differences - SSD, or normalized cross-correlation), often enhanced with optimization techniques like Semi-Global Matching (SGM) to improve accuracy. Code implementation generally involves sliding window comparison and cost aggregation mechanisms.

Optical flow computation estimates pixel motion trajectories across image sequences, with prominent algorithms including Lucas-Kanade (for sparse optical flow) and Farneback (for dense optical flow). These methods operate under the brightness constancy assumption, solving for velocity fields by minimizing grayscale differences before and after pixel motion. Implementation typically involves solving optical flow equations through gradient descent or polynomial expansion techniques.

MATLAB provides convenient toolbox functions such as `stereoDisparity` and `opticalFlowFarneback` for rapid algorithm prototyping and validation. However, challenges like illumination variations, occlusions, and textureless regions require careful consideration. For enhanced robustness, developers can integrate deep learning models such as FlowNet or PWC-Net, which implement convolutional neural networks for end-to-end optical flow estimation.