Reconstruction of Three Dimensional Spatial Clusters Using Monocular Camera

Vision is an increasingly important sensor modality for mobile robots. Visual SLAM (simultaneous localization and mapping) has been used successfully for indoor and outdoor scenarios, where the focus typically lies in localizing the robot and the environment only serves to provide sparse landmarks for a large scale map. Structure from motion (SFM) techniques put emphasis on the actual fine-grained 3D structure of the environment, resulting in a denser more detailed map. For our mobile robot scenario we are interested in detecting objects in such a map. To this end we build on a state-of- the-art SFM technique and use it to build a series of maps, which we stitch together after correcting for scale and finally use the combined map to detect flat surfaces and objects resting on those surfaces.