Probabilistic structure from camera location using straight segments

摘要

A method to determine both the correspondences and the structure from the camera location is presented. Straight image segments are used as features. The location uncertainty is coded using a probabilistic model. The finite length of the image segments is considered, so a more restrictive equation (respect the usage of infinite straight lines) is used, and hence the spurious rejection is improved. The probabilistic modelling derives all the location uncertainty from image error and from camera location error. Thus, the uncertainty is fixed from a physical basis, simplifying the tuning for the matching thresholds. In addition, covariance matrices representing the reconstruction location error are also computed. Experimental results with real images for a trinocular system, and for a sequence of images are presented.