Application of Deep Learning for Drone Obstacle Avoidance and Goal Directed Navigation

Automatic drone flight is a complex task that requires many sensors. Large number of sensors increases the computation power, size and weight of the drone. In this paper, we propose a lightweight automatic flight system that uses only visual sensor for obstacle free navigation in indoor environments. The proposed system is divided into two parts: (1) obstacle avoidance; (2) goal directed navigation. For obstacle avoidance algorithm, we trained a Convolution Neural Network (CNN) to learn potential collision conditions. The drone action is such as to avoid getting in such conditions. Therefore, the dataset used to train the CNN contains various pictures in which the drone gets near or hits obstacles or walls. In the goal navigation system, the drone utilized angle and distance to the AR marker to fly to the target. The performance evaluation is conducted using the AR Drone 2. Experimental results show good performance of drone obstacle avoidance and goal reaching algorithms.