Simultaneous tracking of multiple drones using convolutional neural networks and optical flow

The aim of this paper is to implement a multi-object tracker for detecting, discriminating and following multiple drones on the 2D image plane. The main system consists in a neural network trained for detecting drones on the image plane. Once the drones are identified, their Region Of Interest information are sent to the Kalman filter based tracking system in order to compute position estimates. Each detected drone is given an ID using the Global Nearest Neighbor and Auction Algorithms, then the tracking system employs a set of Kalman filters and each measure is assigned to the respective estimate/filter using the unique ID. In order to reduce the effects of the neural network computation (e.g detection) time, every position estimate is corrected using Optical Flow. Results are evaluated through MOT (Multi-Object Tracking) Metrics, aiming to demonstrate the benefits of implementing Optical Flow algorithms for building a robust multi-tracking system.