Constrained-Covisibility Marginalization For Efficient On-Board Stereo Slam

When targeting embedded applications such as on-board visual localization for small Unmanned Air Vehicles (UAV), available hardware generally becomes a limiting factor. For this reason, the usual strategy is to rely on pure motion integration and/or restricting the size of the map, i.e. performing visual odometry. Moreover, if monocular vision is employed, due to the additional computational cost of stereo vision, this requires dealing with the problem of unknown scale.In this work we discuss how the cost of the tracking task can be reduced without limiting the size of the global map. To do so, the notion of covisibility is strongly used which allows choosing a fixed and optimal set of points to be tracked. Moreover, this work delves into the concept of parallel tracking and mapping and presents some finer parallelization opportunities.Finally, we show how these strategies improve the computational times of a stereo visual SLAM framework called S-PTAM running on-board an embedded computer, close to camera frame-rates and with negligible precision loss.