Real-Time Monocular SLAM With Low Memory Requirements

The localization of a vehicle in an unknown environment is often solved using simultaneous localization and mapping (SLAM) techniques. Many methods have been developed, each requiring a different amount of landmarks (map size), and thus of memory, to work efficiently. Similarly, the required computational time is quite variable from one approach to another. In this paper, we focus on a monocular SLAM problem and propose a new method called MSLAM, which is based on an extended Kalman filter (EKF). The aim is to provide a solution that has low memory and processing time requirements and that can achieve good localization results while benefiting from EKF advantages (i.e., direct access to the covariance matrix, no conversion required for the measures or the state, etc.). To do so, a minimal Cartesian representation (three parameters for three dimensions) is used. However, linearization errors are likely to happen with such a representation. New methods allowing to avoid or hugely decrease the impact of the linearization failures are presented. The first contribution proposed here computes a proper projection of a 3-D uncertainty in the image plane, allowing to track landmarks during longer periods of time. A corrective factor of the Kalman gain is also introduced. It allows to detect wrong updates and correct them, thus reducing the impact of the linearization on the whole system. Our approach is compared with a classic SLAM implementation over different data sets and conditions to illustrate the efficiency of the proposed contributions. The quality of the map built is tested by using it with another vehicle for localization purposes. Finally, a public data set presenting a long trajectory (1.3 km) is also used in order to compare MSLAM with a state-of-the-art monocular EKF-SLAM algorithm, both in terms of accuracy and computational needs.