Autonomous Drone Navigation With Deep Learning

Unmanned Ariel vehicles have become popular in recent years due to the advance in technology,the availability of drone equipment and the low cost of building a drone.As UAVs continue to be popular so is their use case,such that drones are now been used for Ariel surveillance,firefighting,search and rescue and inspection.Over the years,there has been continued research on the development and deployment of drones autonomously.With the help of GPS,sophisticated sensors and complex algorithms such as RGB-D(Red Green Blue-Depth),Light Detection and Ranging(LIDAR),and Simultaneous localization and mapping(SLAM)drone autonomy have achieved a greater height in an outdoor environment.However,the application of this traditional approach in an indoor environment has proven challenging because indoor environments have limited or no GPS signal and the sensors and algorithm are heavyweight and computationally expensive respectively.In this thesis,I proposed an approach that enables a drone using video frame capture by a forward-looking camera to autonomously navigate a drone in a previously unknown indoor environment.The proposed navigation system is achieved using a state-of-the-art two-stream residual neural network(CNN)model.The video frame captured by the forward-looking camera of the drone is fed to the deep learning model to decide on the next command to be issued to the drone.The entire process combine classification and regression task where the model is responsible for producing control command and real distance to a collision.The proposed two-stream residual neural network(CNN)is trained on our custom indoor data sample collected and labelled with the pilot command choice of action and real distance to collision from the external distance sensor mounted on the drone during data sample collection.Leveraging the advantage of skip connection of residual neural network,I trained a deep neural network model without the problem of vanishing gradient.By processing the current frame and its corresponding previous input frame,the model can extract Spatio-temporal features that take into account static appearance and motion information found between frames.The CNN predicted control command and distance to collision are used to modulate the yaw and linear velocity of the drone.Real-time experiment evaluation shows that the proposed approach learns a combined navigation policy of navigating a drone in an unknown indoor environment using just video frames captured by a drone’s forward-looking camera.