Image Rectification and Feature Extraction Based on CUDA for Stereo Vision

When performing traditional binocular vision measurements, the most time-consuming part is the image rectification and feature extraction part, so a CUDA-based image rectification and feature extraction algorithm is proposed, which accomplishes the fast rectification and feature extraction of the images acquired by the binocular camera. A parallelised image rectification algorithm is proposed, which firstly combines the parallelisation feature of CUDA, uses threads for full coverage of the processed image, and uses bilinear interpolation to improve the accuracy of image rectification. The parallelisation of the point feature extraction algorithm is studied, and the algorithm is divided into two parts: connected component labeling and connected component analysis. Based on CUDA technology, the connected component labeling and the statistical analysis of the geometric moment information are quickly completed, and the solution of the centre of mass of the point features is finally completed. Under the same experimental platform, it has been verified through experiments that the improved parallel algorithm, while ensuring accuracy, consumes one tenth of the time of traditional algorithms, achieving a significant increase in computational speed.