Computer Vision Pipeline
Horizon Detection and Scene Segmentation with OpenCV
Completed as a coursework project for the University of Manchester Image Processing module. The system was evaluated against a ground-truth dataset of labelled horizon positions across diverse lighting and terrain conditions.
Outdoor scene understanding requires reliably detecting structural boundaries — particularly the horizon — across images with wildly different lighting, contrast, and weather conditions where fixed-threshold approaches fail.
Implemented the full pipeline in C++ using OpenCV: Gaussian blur for noise reduction, adaptive Otsu thresholding for binarisation, Canny edge detection, and a Hough transform variant to extract dominant linear features. Tuned parameters per image category to improve generalization.
Achieved strong detection accuracy across the evaluation dataset, with the pipeline handling edge cases including overcast skies, treelines, and partial occlusion.
