Practically Classifying Unmanned Aerial Vehicles Sound Using Convolutional Neural Networks

Abstract

In this work, we analyze the effectiveness of a simple neural network for the task of determining, by sound, if small unmanned vehicles are carrying potentially harmful payloads The goal of this work is to contribute to a real-time UAV detection system that requires a means of assessing threat level of incoming vehicles whose positions are determined by other sensors. Further, we operated under a minimal cost constraints to enable eventual adoption at scale by law enforcement agencies. Our system classifies payload carrying vs. non-payload carrying DJI Phantom II UAVs by presenting sound spectrum data to a simple Convolutional Neural Networks (CNN). These networks, along with a simple voting system, provided a 99.92% recognition rate for this problem without a need to violate our minimal cost constraint.