Gradient Operators for the Determination of Optical Flow

Gradient operators for the determination of optical flow in images are designed using classical linear regression methods. Two general approaches are presented: the first method assumes that the linear models in each spatial and temporal dimension are separable; the second (and less-commonly used) method does not rely on this assumption. The second method allows complex shapes and modes of motion to be captured. The methods are optimized by adjusting filter latency, model order and window size. Performance is compared using synthetic data. The second method is shown to outperform the first method.