Stereo Vision Based Object Detection for Mobile Robots (Stereo Visione per il Riconoscimento di Oggetti nella Robotica Mobile)

This method proposed in the paper Tao et al. (2008) is an improved block matching approach for fast stereo correspondence. The matching criterion is the Sum of Absolute Difference (SAD) and the approach tries to progressively confine the matching operation during the searching process.
As already pointed, the block matching method is the most popular local method because of its simplicity for the implementation. To find the stereo correspondence it estimates the point disparity in the left image and a block surrounding this point; in the second step it will check for the closest matched block within a search range in the right image. The displacement between the reference block and the closest matched box depicts the disparity of the point just evaluated. Therefore the correspondence problem results essentially in a search problem and the standard search method is a wasteful process where the matching criterion is calculated for all pixels at all possible search positions.
So, as easily understandable, it is not necessary to calculate the accuracy for the matching error for the entire region because, for instance, it’s possible to reject a matching block whenever the measured partial error between the candidate block and the reference block is higher than a previously found error.
This last idea is therefore used to improve the basic stereo matching algorithm. Once the disparity map is generated, the algorithmuses a filter to try to locate a correlation peak between pixel and to performa sub-pixel accuracy.
The OpenCV library provides the implementation of the method which was then used to test the BM-Stereo functionalities on our training set of images. The choice of the parameters to be used is an essential point and it has to be done accordingly to the quality of the used cameras and the environment in which they are suppose to work. [...]