Computation of optical flow using a neural network

Publisher Website
Google Scholar
Abstract
A method for computing optical flow using a neural network is presented. Usually, the measurement primitives used for computing optical flow from successive image frames are the image-intensity values and their spatial and temporal derivatives, and tokens such as edges, corners, and linear features. Conventional methods based on such primitives suffer from edge sparsity, noise distortion, or sensitivity to rotation. The authors first fit a 2-D polynomial to find a smooth continuous image-intensity function in a window and estimate the subpixel intensity values and their principal curvatures. Under the local rigidity assumption and smoothness constraints, a neural network is then used to implement the computing procedure based on the estimated intensity values and their principal curvatures. Owing to the dense measured primitives, a dense optical flow with subpixel accuracy is obtained with only a few iterations. Since intensity values and their principle curvatures are rotation-invariant, this method can detect both rotating and translating objects in the scene. Experimental results using synthetic image sequences demonstrate the efficacy of the method.

This publication has 14 references indexed in Scilit: