A neural scheme for optical flow computation based on Gabor filters and generalized gradient method

Abstract A neural scheme for computing optical flow from the intensity image and its time derivative is presented. The receptive fields of the cells in the magno division of the primate visual cortex are known to be orientation selective and bandpass. Modelled as Gabor filters and differential Gabor filters, the magno cells can be organized as motion sensors for extracting generalized spatiotemporal gradients from the intensity images and their time derivatives. The optical flow is computed from the generalized spatiotemporal gradients with a neural circuit for finding the least-squared-error. The neural scheme avoids some critical shortcomings of current neural schemes for optical flow computation, particularly the gradient schemes based on the discrete spatial differentiations and the Fourier scheme based on the spectral analysis on the image sequences. Our computational tests on synthetic and natural image data show that our scheme is accurate to natural scenes.