An underwater binocular vision positioning method based on back-propagation neural networks

Abstract

In view of the existing underwater sensors cannot obtain a relative positional relationship between the target and the underwater vehicle, as well as, target positioning based on binocular vision method which has systematic errors needs more accurate calibration result. An approach for vision of a recognized underwater target is proposed in this paper. The first step is to process the original image with image enhancement and filtering, identify the target image based on HSV threshold. In the next step for target positioning, a three dimensional (3D) re-construction method based on block matching after binocular vision calibration and a 3D re-construction method with learning neural network are studied. Aiming at the problems of slow convergence rate and local optimum of traditional BP networks, several improvements of traditional BP networks are proposed. Finally, the experiment proves that the method can obtain high precision result and good real-time performance under sufficient stylebook data.