Stereo-vision is a technique in image processing used for depth estimation, utilizing two calibrated cameras positioned at a fixed distance apart. This repository provides sample code for testing a stereo-vision setup using OpenCV (CV2).
Stereo-vision mimics human binocular vision, enabling depth perception by comparing the slightly different images received from each eye. Humans possess two eyes that allow them to gauge the depth of a scene. This capability arises from the fixed distance between the eyes, which capture images from different angles, resulting in a disparity between the images. The human brain interprets this disparity to understand depth, as images are essentially 2D representations of a 3D environment.
In a stereo-vision system, two cameras are employed at a fixed distance. Because the images captured by each camera originate from slightly different angles, the positions of objects in each image differ slightly, creating what is known as disparity. Various matching methods, such as template matching, can be used to calculate this disparity. It can then be interpreted as distance using the following formula:
where
The calibration of the cameras plays a crucial role in stereo-vision. Therefore, the quality of depth estimation is directly related to the quality of the calibration. The lenses of the cameras often cause distortion in the captured images. Additionally, the cameras need to be calibrated in relation to each other. The relative rotations of the cameras can be explained using a homography matrix, which can then be provided to the stereo-vision algorithm.
To calibrate your stereo-vision setup, I highly recommend visiting this repository: python_stereo_camera_calibrate.
It provides an easy and automated calibration algorithm.
The following image shows the captured frames from a stereo-vision setup taken simultaneously. Notably, the images appear to be identical.
The dataset for the stereo-vision tests executed here is from the FGI Masala Stereo-Visual-Inertial Dataset, which is available here.
The following images are the results of executing the stereo-vision algorithm on images from two calibrated cameras that were captured simultaneously.
It can be seen that the parts of the scene in the left-side images (gray-scale) that are closer to the cameras stand out in the output image (color-map).
