Intelligent Learning Algorithm for Depth Detection Applied to a Humanoid Vision System

摘要

We can utilize stereo-pair images obtained from two cameras to compute world coordinate points by using the principle of triangulation. However, there are some restrictions from cameras and some initial parameters need to be obtained in the experiment, by applying this method to get world coordinates. In this paper, a type of depth detection training system is proposed that artificial neural networks be used to train the system such that the need for the initial parameters of cameras are eliminated. The training set for my neural network consists of a variety of target points in stereo-pair and the corresponding world coordinates of points. Through the error analysis, we can know the neural network depth detection system' accuracy is far greater than the traditional depth detection method, and the method is enough accurate that can be widely used in stereo vision system.