Abstract: The accuracy of spatial calibration in Stereo Particle Image Velocimetry (SPIV) has a considerable influence on the accuracy of velocity measurement. To study the ability of various calibration models to handle input errors, we define a dimensionless parameter, namely, the error attenuation coefficient, to evaluate the response of spatial calibration models to input errors. Based on this error attenuation coefficient, the error propagation characteristics of conventional spatial calibration models, including the polynomial model and the camera pinhole model, can be quantitatively evaluated. A neural network-based space calibration model is then developed. Unlike conventional calibration models, this new model is naturally adaptive to multiple-camera joint calibration, thus suitable for SPIV. Synthetic experiments demonstrate the ability of this neural network model to suppress the propagation of the input error in a large measurement parameter space, which is not possessed by the polynomial model or the pinhole model. Additionally, it outperforms traditional models in the scenario of high optical distortion. Therefore, this neural network model might be an ideal candidate for the spatial calibration of SPIV. Finally, it is confirmed in the experiment that the error of neural network calibration model is only a quarter of that of traditional models.

Key words: SPIV, neural networks, camera calibration, machine vision, spatial positioning