Undelayed landmarks initialization for monocular SLAM

We address the problem of landmark initialization in monocular simultaneous localization and mapping (SLAM). The depth dimension is not observable from one monocular measurement, and several observations are required from different vantage points exhibiting sufficient parallax. This makes initialization difficult. Early solutions to the problem performed a parallel task to determine this depth before initializing, dealing to what we name delayed methods. We show that these methods cannot exploit the angular information provided by low parallax landmarks. We propose the design of undelayed methods, where the landmarks are immediately initialized without the depth information. Low parallax landmarks include those close to the motion axis and those at remote distances, potentially at infinity. Suitable depth parameterizations are required for undelayed methods to work with distant landmarks. We introduce a Gaussian mixture parameterization, with its terms geometrically distributed along the landmark’s visual axis. We obtain a scale invariant representation that scales nicely to very long distances. To avoid the exponential explosion of the problem size inherent to this multiGaussian scheme, we develop special mapping and updating procedures that exhibit a linear growth of the SLAM state vector. The methods are tested and validated via simulation and real experiments, in both indoor and outdoor scenarios, and with and without the aid of odometry data.