Scene Segmentation by Velocity Measurements Obtained with a Cross-Shaped Template

Abstract

This paper presents a method of segmenting a scene into its basic objects on the basis of motion. The velocities of points are approximated by associating them to sets of reference edges in the scene. These measurements are then used to group the points, A basic tenant of this work is that all points of a single object have the same velocity values. Accordingly, points with the same velocity measurements are grouped together. Since the motion of an object is independent of its other visual characteristics, whole objects (not surfaces or edges) are initially included in one segment. The program successfully grouped points from internally generated scenes with 100% accuracy. Points from real test scenes were grouped correctly 90% of the time on the average. The poorest result was 42% accuracy The best was 100%.