Real-Time Drone Anti-Collision Avoidance Systems: an Edge Artificial Intelligence Application

Collision Avoidance (CA) is one of the most challenging and fundamental problems in the Unmanned Aerial Vehicle (UAV) ecosystem. The unmanned version of Airborne Collisions Avoidance Systems (ACAS-Xu) detects collision threats and informs pilots of avoidance maneuvers to perform. It makes decisions based on an optimized cost tables. However, these tables are too large and heavy (~5GB) to be integrated on the current certified avionic hardware. In this paper, we present a machine learning (ML) based approach to reduce the memory footprint of the cost tables by approximating them using a deep neural network (DNN). Training results show that the proposed approach reduces the memory footprint drastically by a factor of 500 from 5GB to 10MB while maintaining the same level of safety. Additionally, we embedded the DNN based approach on an edge artificial intelligence (AI) device and demonstrated its ability to operate on limited hardware with low processing time.