3DoF Bearing-Only SLAM
Bearing-Only SLAM is particularly suited to match the requirements of service robots since these can be used with cheap sensors like omnidirectional cameras. As long as no calibrated systems are needed, omnidirectional cameras are cheap and small and thus suitable for service robots. Although omnidirectional cameras provide feature rich information on the surrounding of the robot with high update rates, they do not provide range information.
Thus, one has to modify the sensor models of the well-known SLAM approaches such that observation angles of landmarks are suffcient to generate pose estimates. The problem is that one needs several observations of the same landmark from different poses to intersect the lines of sights. Thus, one has to solve the so-called problem of a delayed initialization of a landmark [Bai03]. The problem results from the fact that the estimates of the observation poses of not yet initialized landmarks have to be corrected with each re-observation of an already known landmark. Thus, one extends the state vector such that it not only contains the robot pose and the initialized landmark poses but also the observation poses of not yet initialized landmarks.
Now these observation poses are consistently updated by the SLAM mechanism. Since the measurements are relative to the observation poses and since these are updated consistently, one can consistently transform angular observations into line of sights in the global frame of reference even if the measurements are from different points in time. Thus, a deferred but consistent evaluation of measurements is possible. Measurements of a possible landmark can then be accumulated over time until suffcient information is available for a reliable landmark initialization.