Design and Evaluation of a Long Baseline Stereo Vision System

Sammanfattning

Stereo vision systems can estimate the distance to objects from a pair of cameras, but their accuracy decreases rapidly as objects move further away. This makes it difficult to use them for measuring the size of objects at long distances. The aim of this project was to design and evaluate a stereo vision system that can measure the dimensions of objects at a distance of 20 metres with a relative depth error below 2.5%. A custom stereo rig was built that allows the distance between the two cameras to be changed, so that several configurations could be tested. For each configuration, the cameras were calibrated and images were captured of a known reference object at different distances. The same images were then processed by two methods: a classical algorithm and a modern deep learning model. The results were compared by measuring the width, height, and depth of the reference object in the resulting 3D point clouds. The deep learning model met the accuracy requirement at 20 metres for almost all tested configurations, while the classical algorithm met the requirement only for a small number of them. The findings show that combining a wide-baseline stereo rig with a modern deep learning depth estimator is an effective and affordable way to measure objects at long distances.