← Back to Research
Research Note · Computer Vision
Computer Vision Experiments
Published: Jun 2026•5 min read•Author: Chetraj Jaishi
OpenCVPythonOptical FlowCalibration
Abstract: Hands-on experimentation with image processing pipelines, feature descriptors, camera calibrations, and optical flow.
Before feeding images blindly into neural networks, understanding classical computer vision fundamentals—Sobel gradients, Hough transforms, Harris corner detection, and pinhole camera models—provides crucial intuition.
In this series of experiments, we implement Lucas-Kanade optical flow tracking for velocity estimation, compute intrinsic camera calibration matrices using planar checkerboards, and explore epipolar geometry for stereo disparity estimation.
These classical methods remain indispensable in robotics for low-latency visual odometry, where running heavy deep learning models for every single state estimate is computationally prohibitive.
Key Research Takeaways
- Camera calibration and lens distortion correction (radial and tangential) are prerequisite steps for any accurate metric calculation.
- Optical flow provides immediate velocity vectors without requiring training data or GPU acceleration.
- Combining classical feature tracking with modern deep learning representations yields the most robust real-world vision systems.