Drone Collision Detection and Classification using Proprioceptive Data

This paper proposes an approach for collision detection and obstacle classification based on the physical interaction between a drone and the environment. Our method does not require any special collision-detection sensor, since it uses only proprioceptive data (accelerations). Time-series classification of acceleration data is carried out by a deep neural network, which uses acceleration time-series to classify the drone’s flight mode into three classes: no collision, collision with a soft obstacle, and collision with a hard obstacle. Experimental results showed that the classification achieved accuracy of 98.7% on the testing set. Also, the method was used to perform on-line flight state predictions, when data is applied continuously into the neural network’s input, thus showing the potential of this method for collision detection and classification in several applications.