Stereoscopic vision calibration for three-dimensional tracking velocimetry based on artificial neural networks

Here, the physical and mathematical model is briefly described first, on which the photogrammetric calibration procedure of our Stereoscopic Tracking Velocimetry (STV) system is based. A new hybrid calibration approach is then introduced, which incorporates the use of artificial neural networks. The concept is to improve the performances of conventional calibration techniques of stereoscopic vision. In order to evaluate the quality of the hybrid calibration approach, calibration error is defined for the use of a camera. Our experimental investigation shows that the accuracy in predicting the object frame coordinates has been improved by 30 percents twhen the hybrid calibration is employed, as compared with the case when only the previous conventional physical and mathematical model is directly applied. It appears that the new idea of using artificial neural networks together with a physical and mathematical model of a system can improve the overall performance of the system. The hybrid method can also be applicable to other general areas in machine vision.