Efficient on-board Stereo SLAM through constrained-covisibility strategies

Abstract

Visual SLAM is a computationally expensive task, with a complexity that grows unbounded as the size of the explored area increases. This becomes an issue when targeting embedded applications such as on-board localization on Micro Aerial Vehicles (MAVs), where real-time execution is mandatory and computational resources are a limiting factor. The herein proposed method introduces a covisibility-graph based map representation which allows a visual SLAM system to execute with a complexity that does not depend on the size of the map. The proposed structure allows to efficiently select locally relevant portions of the map to be optimized in such a way that the results resemble performing a full optimization on the whole trajectory. We build on S-PTAM (Stereo Parallel Tracking and Mapping), yielding an accurate and robust stereo SLAM system capable to work in real-time under limited hardware constraints such as those present in MAVs. The developed SLAM system in assessed using the EuRoC dataset. Results show that covisibility-graph based map culling allows the SLAM system to run in real-time even on a low-resource embedded computer. The impact of each SLAM task on the overall system performance is analyzed in detail and the SLAM system is compared with state-of-the-art methods to validate the presented approach.