Towards 3D visual SLAM for an autonomous quadcopter running on ROS

In a context where multi-copters gain in attractiveness each day to solve novel challenges or replace older technologies, the need for autonomous behaviour is being pointed out as a key feature for unlocking lots of applications. This master thesis is conducted at UCL for trying to understand the current challenges to unlock multi-copter autonomy. We point out that the lack of robustness of all current state-of-the-art implementations which are based only upon on-board sensors, leads, as a common flaw, to a lack of accuracy of the pose estimation. Indeed, for GPS-denied environments such as indoors, there exist poor absolute references to build an accurate and robust belief of the position and orientation of the drone. Unfortunately, indoor situations need precisely the best pose estimation since they are confined places not allowing too large errors while executing movement. We propose an implementation of the direct keyframe-based visual SLAM approach based on features detected in the images of the video-stream of an embedded camera. We show that the results obtained on a real low cost quadcopter, the Parrot AR.Drone, improve substantially the pose estimation, notably by canceling the drift on sensor readings. However, this technique presents strong limitations when confronted to untextured environments, due to the lack of reference keypoints detected in the images. To palliate this, we explore the state of the literature to propose some promising advances to replace or enhance this approach, such as replacing conventional cameras by RGB-D (Red Green Blue - Depth) cameras or DVS (Dynamic Vision Sensors), or replacing computationally costly software parts by a hardware implementation, or better fusing the available sensors informations.