Multi-camera VSLAM : from former information losses to self-calibration

Visual SLAM is, in recent years, a very active research area, the result of activities in the convergence of the robotics and computer vision communities. We present an overview of techniques, from classical filtering to bundle adjustment solutions, for both monocular and stereo (or multicamera) systems, and emphasize that classical SLAM solutions have been discarding precious sensory information. In particular, the ability of vision to sense objects at infinity should be exploited at its maximum because it is precisely these remote objects that will provide long-term, stable angular references (in the way a compass would do). Monocular SLAM systems have already solved this issue, but stereo and multicamera systems have not. We propose in these cases to use monocular SLAM algorithms using fusion to incorporate all the information. Numerous advantages like desynchronization of the sensors firing, the possibility of using several unequal cameras, or selfcalibration, will naturally arise. We develop a particular method for extrinsically decalibrated stereo systems to illustrate the proposed ideas. We evaluate the method with a real indoor experiment, and highlight and discuss both its assets and drawbacks.