Reconstruction of 3D environments is a problem thatudhas been widely addressed in the literature. While manyudapproaches exist to perform reconstruction, few of themudtake an active role in deciding where the next observationsudshould come from. Furthermore, the problem of travellingudfrom the camera’s current position to the next, known asudpathplanning, usually focuses on minimising path length.udThis approach is ill-suited for reconstruction applications,udwhere learning about the environment is more valuable thanudspeed of traversal.udWe present a novel Scenic Route Planner that selectsudpaths which maximise information gain, both in terms ofudtotal map coverage and reconstruction accuracy. We alsoudintroduce a new type of collaborative behaviour into theudplanning stage called opportunistic collaboration, whichudallows sensors to switch between acting as independentudStructure from Motion (SfM) agents or as a variable baselineudstereo pair.udWe show that Scenic Planning enables similar performanceudto state-of-the-art batch approaches using less thanud0.00027% of the possible stereo pairs (3% of the views).udComparison against length-based pathplanning approachesudshow that our approach produces more complete and moreudaccurate maps with fewer frames. Finally, we demonstrateudthe Scenic Pathplanner’s ability to generalise to live scenariosudby mounting cameras on autonomous ground-basedudsensor platforms and exploring an environment.
展开▼
