LINEAR CONSTRAINTS IN TWO-VIEW MULTIPLE HOMOGRAPHY ESTIMATION OF UNCALIBRATED SCENES

摘要

In this contribution, we present a method for direct linear estimation of multiple homographies and optimization of estimation results by enforcing topological constraints by means of connected components. The applications are the detection of independent moving objects seen by a moving observer and sparse reconstruction of the scene. Security and surveillance applications require day-and-night capability. Therefore thermal cameras are the preferred sensors. These sensors violate primary constraints for the estimation of the optical flow. The motion of the sensor is compensated by image stabilization, which allows us to utilize results from the extensive field of change detection, for example, background subtraction and temporal differences. We assume that the scene can be approximated by a composition of several planes. Each plane induces an independent homography for each image pair. A set of homographies for one image pair is called consistent if and only if all homographies refer to the same relative orientation but even though they may refer to different 3d-planes. We show that linear constraints can be induced to homography estimation to ensure consistency within a pair of views. The fundamental matrix allows us to formulate a solution independent of the calibration of the camera. Given a set of consistent homographies and a corresponding image segmentation for an image pair it is possible to improve the image stabilization by local warping instead of warping with one single homography. This approach is explored by experiments on disparity map estimation based on the homography estimation on the Middlebury-Stereo Benchmark dataset. Additionally, the robustness of the algorithm is explored using a synthetic scene. Finally we show first results of motion detection based on this method combined with temporal differences.