A multi-stage visual odometry system and failure detection mechanism for a monocular visual SLAM implementation in monotonously textured underwater scenes

摘要

A multi stage visual odometry system with a failure detection mechanism is developed to measure movement and correct unreliable estimations in poorly textured underwater environments. The system has only one camera as the sensor, and is composed of a primary movement estimation based of features presented in the scene and two backup system. The first backup is based on phase correlation to estimate the AUV movement. The second backup movement estimation mechanism is based on a high order prediction of the movement using the last good movement estimations. Failure detection mechanism are constructed to detect unreliable measurements in each of the stages and appropriate responses are developed in all cases. A full SLAM implementation is developed and a real case study is presented in which the good properties of the system are shown.