Toward Multi-Stage Decoupled Visual SLAM System

SLAM is defined as simultaneous estimation of mobile robot pose and structure of the surrounding environment. Currently, there is a much interest in Visual SLAM, SLAM with a camera as main sensor, because the camera is an ubiquitous and affordable sensor. Camera measurements formed by perspective projection is highly nonlinear with respect to estimated states, leading to complicated nonlinear estimation problem. In this paper, a novel system is proposed that divides the problem into two parts: local and global motion estimation. This division leads to a simple linear estimation system. In the first stage, local motion parameters (acceleration, velocity, angular acceleration and orientation) are estimated in robot local frame. Robot position and the scene map are then estimated in the second stage in global frame as global motion parameters. Map is updated at each camera frame and is represented in a relative way to decouple robot pose from map structure estimation. The new system simplified the map correction to a linear optimization problem. Simulation results showed that the proposed system converges and yields accurate results.