Camera calibration is the base of the machine vision based the autonomous navigation of the agricultural wheeled-mobile robots. There are the complex nonlinear relationship between the actual position points and the matched image points. Therefore the camera parameters have to be calculated by a precise imaging model. The more precise the imaging model requires, the more complicated the calibration becomes. It was proved that some traditional calibration methods, such as the method of Zhengyou Zhang, were inconvenient and their accuracy were also low. In this paper, according to the characteristic of the BP neural network, which can express any nonlinear relationship between inputs and outputs, a new method based on the BP neural network was applied to calibrate the vision system of an agricultural wheeled-mobile robot. The experimental results showed that this method was feasible and accurate.