Measurement and integration of 3-D structures by tracking edge lines

This article describes techniques for dynamically modeling the 2-D appearance and 3-D geometry of a scene by integrating information from a moving camera. These techniques are illustrated by the design of a system that constructs a geometric description of a scene from the motion of a camera mounted on a robot arm.A framework for dynamic world modeling is described. The framwork presents the fusion of perceptual information as a cyclic process composed of three phases: Predict, Match, and Update. A set of mathematical tools are presented for each of these phases. The use of these tools is illustrated by the design of a system for tracking edge lines in image coordinates and inferring the 3-D position from a known camera motion.The movement of edge-lines in a sequence of images is measured by tracking to maintain an image-plane description of movement. This description is composed of a list of edge-segments represented as a parametric primitive. Each parameter is composed of an estimated value, a temporal derivative, and a covariance matrix. Line-segment parameters are updated using a Kalman filter. The correspondence between observed and predicted segments is determined by a nearest-neighbor matching algorithm using distance between parameters normalized by covariance. It is observed that constraining the acceleration of edge-lines between frames permits the use of a very simple matching algorihtm, thus yielding a very short cycle time.Three-dimensional structure is computed using the correspondence provided by the 2-D segment tracking process. Fusion of 3-D data from different view points provides an accurate representation of the geometry of objects in the scene. An extended Kalman filter is applied to the inference of the 3-D position and orientation parameters of 2-D segments. This process demonstrates that 2-D tracking provides the information for an inexpensive technique for estimating 3-D shape from motion.Results from several image sequences taken from a camera mounted on a robot arm are presented to illustrate the reliability and precision of the technique.