Real-time visual-inertial slam for multirotor platform

Abstract

This thesis presents the theoretical and practical development of a framework that enables small-scale aerial vehicles to autonomously explore and map unknown environments. The increased demands of all-terrain navigational capability and full 6 Degrees-of- Freedom (DOF) solutions have driven the usage of Inertial Measurement Unit (IMU) and related Inertial Navigation System (INS). Unfortunately, the INS is a highly unstable system, containing three integrator loops coupled with the presence of gravity, causing errors to accumulate cubically as time increases. To constrain these errors effectively, a reliable and constant stream of aiding signals is crucial for the success of inertial applications. Therefore to utilise the aiding information at higher rates, this work proposes a novel visual non-holonomic constraint. These relative pose constraints are estimated from the consecutive monocular image sequence and capable of providing the ego-motion estimates at frame-rate to inertial sensor. The basic concept is the tangential velocity components of INS along the direction of visual-flow motion should be zero, that is similar to the non-holonomic constraint (or non-skidding assumption) in the wheeled robots [1, 2]. The contributions of this thesis are as follows. Firstly, the monocular SLAM using the concept of Visual-Inertial non-holonomic constraint is investigated and analysed. The monocular SLAM generally estimates the pose of the vehicle and landmarks jointly. The fundamental problem with these approaches is that the computational requirements scale quadratically in the number of landmarks. Therefore, this limits the number of visual features to only a few hundred and small-scale environment. We fused the visual non-holonomic constraints (using hundreds of features) at high update rates with an inertial sensor. In this way, by using these constraints the on-board IMU/INS can be calibrated and corrected accurately without maintaining a large number of features. Secondly, a depth dropout problem in RGB-D (Kinect or Stereo) SLAM is addressed and researched. Depth data from the RGB-D [5, 6] can be easily unavailable or partially available due to limited range or sunlight interference. The unavailability of reliable depth data makes the existing RGB-D approaches inapplicable for many outdoor environments. We resolved the depth dropout issue by providing either full 6DOF (rotation and translation RGB-D constraints) or partial 5DOF visual non-holonomic constraints to the inertial sensor. Thirdly, the convexity structure of an Inertial-SLAM was investigated by analysing the number of local minima, which is important to understand the pseudo-optimality of the SLAM system. This sheds light on solving the non-linear SLAM problem without resorting to the exhaustive search techniques in global optimisation. Lastly, the proposed methods are implemented in an embedded computer on the hexarotor platform and its real-time performance is demonstrated in challenging outdoor environments. A customised, low-cost IMU is developed in-house and the real-time on-board implementation of the proposed framework is performed. A demonstration of the developed framework is evaluated with hovering control. For the purpose of demonstration, an attitude and position controller is also developed through an open-source autopilot system. The accuracy and performance are demonstrated against the precise ground truth data Show more