Analysis of video image sequences using point and line correspondences

摘要

In this research, we investigate the problem of analyzing time-varying imagery using a feature-based approach. We assume a scenario where the imaged objects remain stationary while the camera moves. The goal is to compute the structure of the imaged objects and the motion of the camera from a sequence of video images. The proposed method exploits the principle of the invariance of rigid configurations during motion. Using the rigidity constraints, we specify equations based on distance and angular invariance to compute the structural parameters of the imaged 3D objects independently of the camera's motion. Once the structural parameters are recovered, the motion parameters can be computed. The strength of our approach is in the decomposition of the computations of structure and motion, and in the use of point and line correspondences simultaneously. The main advantage of using both points and lines is that our approach is not limited to objects whose images have only a particular type of feature in abundance. In this research, we consider the special case of four points and one line, which are the minimum feature sets required for computing the structure and motion parameters. We then consider additional features in an overdetermined system of equations to improve the reliability and accuracy of the computation. We present computer simulation results, as well as results from real image sequences, to demonstrate the algorithm's validity.