Aiming at the problems of long feature point matching time, high false matching rate and low binocular vision measurement accuracy of traditional algorithms, a binocular vision measurement method based on two-dimensional entropy and low-dimensional descriptor is proposed. Firstly, the two-dimensional entropy of the image is used to screen the feature points, filter some useless feature points, and ensure the stability of the feature points. Then, a low-dimensional SIFT feature descriptor with multiple gradient directions is constructed to improve the discriminative ability of feature points and to obtain more gradient information of seed points, while using the mahalanobis distance as a similarity metric for feature point matching and eliminating false matches with the random sampling consistency RANSAC algorithm to optimize the matching accuracy and to reduce the complexity of the algorithm. Finally the sub-pixel coordinates of the feature points are obtained by binary quadratic surface fitting, and the spatial 3D coordinates are obtained by triangulation to calculate the relevant dimensions of the measured object. Taking the continuous casting slab model as the measurement object, the experimental results show that the average relative error of the measurement is 0.41%, which is 1.45% and 0.72% lower than that of the SIFT algorithm and ORB algorithm, respectively, and meets the requirements of the measurement accuracy. The correct rate of the feature point matching improves by 20.94%, 18.19% and 11.38% compared with that of the SIFT, BRISK, and ORB algorithms, and the time taken for feature point matching reduces by 57.48% compared with SIFT, which verifies the accuracy and efficiency of the algorithm.