Dimensionless Monocular SLAM

It has recently been demonstrated that the fundamental computer vision problem of structure from motion with a single camera can be tackled using the sequential, probabilistic methodology of monocular SLAM (Simultaneous Localisation and Mapping). A key part of this approach is to use the priors available on camera motion and scene structure to aid robust real-time tracking and ultimately enable metric motion and scene reconstruction. In particular, a scene object of known size is normally used to initialise tracking. In this paper we show that real-time monocular SLAM can be initialised with no prior knowledge of scene objects within the context of a powerful new dimensionless understanding and parameterisation of the problem. When a single camera moves through a scene with no extra sensing, the scale of the whole motion and map is not observable, but we show that up-to-scale quantities can be robustly estimated. Further we describe how the monocular SLAM state vector can be partitioned into two parts: a dimensionless part, representing up-to-scale scene and camera motion geometry, and an extra metric parameter representing scale. The dimensionless parameterisation permits tuning of the probabilistic SLAM filter in terms of image values, without any assumptions about scene scale, but scale information can be put back into the estimation if it becomes available. Experimental results with real image sequences showing SLAM without an initialisation object, different image tuning examples and scenes with the same underlying dimensionless geometry are presented.