An Integrated Neural and Algorithmic System for Optical Flow Computation

Motion detection plays a central role in several visual environments: knowledge of object velocities and trajectories is fundamental in scene interpretation and segmentation. This task appears a simple problem, but detecting moving objects is very difficult, in fact this is a problem that cannot be considered completely solved today [1] [2] [3].

In this paper we present a novel method that uses two different approaches: a “neural” one and an algorithmic one. In fact, a Multilayer Perceptron is used in the first stage, in order to detect some motion areas in the scene [5] [6]; a matching algorithm is then used to obtain a sparse optical flow and to compute the epipolar geometry of the moving camera [7] [8]; and, finally, a refinement algorithm is used to produce a denser optical flow field. Thus this method can extract features automatically from moving objects in a scene discarding stationary ones. This approach seems to be very useful for tracking and motion segmentation.

This work was developed in the context of JACOB project, to achieve the automatic retrieval of images based on motion [9].