Shape from pairwise silhouettes for plan-view map generation

摘要

A plan-view map is a representation mechanism especially appropriated for people detection and tracking. It allows to represent volumetric information in a very compact and efficient way so that very effective detection and tracking algorithms can be applied using it. In spite of their advantages, plan-view maps have mainly been used with stereo cameras but not with monocular ones. The reason why is that many of the three-dimensional reconstruction algorithms using monocular vision impose strong conditions (such as controlled lightning, lack of occlusion, etc.) thus impeding their applicability in realistic environments.This paper presents two main novelties. First, a three-dimensional reconstruction algorithm that is especially appropriate for people detection and tracking because of its robustness to errors in the background subtraction. Second, a new technique for creating plan-view maps from volumetric reconstructions that reduces the amount of false positives in the occluded regions. We show experimentally that these techniques produce plan-view maps that can be reliably employed for tracking purposes. The proposed methods are compared with the traditional SfS obtaining relevant improvements.