Combining Deep Learning And Rgbd Slam For Monocular Indoor Autonomous Flight

We present a system that uses deep learning and visual SLAM for autonomous flight in indoor environments. In this spirit, we use a state-of-the-art CNN architecture to obtain depth estimates, on a frame-to-frame basis, of images obtained from the drone's onboard camera, and use them in a visual SLAM system to obtain both camera pose estimates with a metric that is further passed to a PID controller, responsible for the autonomous flight. However, because depth estimation and visual SLAM system are computationally intensive tasks, the processing is carried out off-board on a ground control station that receives online imagery and inertial data transmitted by the drone via a WiFi channel during the flight mission. Further, the metric pose estimates are used by the PID controller that communicates back to the vehicle with the caveat that synchronisation issues may arise in between the frame reception and the pose estimation output, typically with the frame reception running at 30 Hz, and the pose estimation at 15Hz. As a consequence, the controller may also exhibit a delay in the control loop, provoking a flight off-track the trajectory set by the way-points. To mitigate this, we implemented a stochastic filter that estimates velocity and acceleration of the vehicle to predict pose estimates in those frames where no pose estimate is available yet, and when available, to compensate for the communication delay. We have evaluated the use of this methodology for indoor autonomous flight with promising results.