Multi-robot cooperative stereo for outdoor scenarios

In this paper, we propose a cooperative perception framework for multi-robot real-time 3D high dynamic target estimation in outdoor scenarios based on monocular camera available on each robot. The relative position and orientation between robots establishes a flexible and dynamic stereo baseline. Overlap views subject to geometric constraints emerged from the stereo formulation, which allowed us to obtain a decentralized cooperative perception layer. Epipolar constraints related to the global frame are applied both in image feature matching and to feature searching and detection optimization in the image processing of robots with low computational capabilities. In contrast to classic stereo, the proposed framework considers all sources of uncertainty (in localization, attitude and image detection from both robots) in the determination of the objects best 3D localization and its uncertainty. The proposed framework can be later integrated in a decentralized data fusion (DDF) multi-target tracking approach where it can contribute to reduce rumor propagation data association and track initialization issues. We demonstrate the advantages of this approach in real outdoor scenario. This is done by comparing a stereo rigid baseline standalone target tracking with the proposed multi-robot cooperative stereo between a micro aerial vehicle (MAV) and an autonomous ground vehicle (AGV).