A Vision System for Depth Perception that Uses Inertial Sensing and Motion Parallax

Abstract

There are several ways for a vision system to obtain depth information from a complex visual scene. Some extract absolute distance information and others only relative information from the scene. Commonly used methods that give absolute distance measurements are triangulation, suitable for medium to far distances, and binocular disparity used in stereo vision, which is a robust method for determining distance information at close range. Other cues that are commonly used, but that only give relative distance information, are scale, perspective, precedence of non-occlusion etc. There is however another important cue for distance, or depth perception, not as commonly used as stereo vision, and that can reveal depth information from a monocular motion field. It is referred to as structure from motion or kinetic depth effect when the object that is observed is moving. When the observer is moving it is referred to as motion parallax (Mallot 1998, p 201). The former gives rise to local depth information and the latter can determine distances to objects. There is however an important requirement for determining absolute distances to objects, i.e. that the motion of the observer needs to be known. The motion of the observer will in our case correspond to the measured inertial change of the observer. The distance to an object point can then be obtained by augmenting the additional inertial cue, i.e. the observer motion, with the perceived motion of any object point present in the scene (Huster 2003, Karri et al. 2005).