Stone Soup: a Route to Impact for Surveillance-relevant Computer Science

25^th November 2020, 11:00
Simon Maskell and Lyudmil Vladimirov

Abstract

Stone Soup is an open-source python software library that, prior to beta-release, was developed by a team spanning the UK, US, Canadian and Australian governments and the University of Liverpool. Stone Soup collates algorithms for processing data from sensors (eg cameras, radars, sonar, AIS receivers etc) to detect, track and classify objects in sequential streams of data. While some of the algorithms used (eg YOLO and kd-trees) simply involve interfacing to pre-existing libraries, the focus is on algorithms for which open-literature modular implementations are not otherwise available. Stone Soup therefore already includes baseline algorithms like Kalman filters as well as more sophisticated algorithms such as Joint Probabilistic Data Association (an approach based on matrix permanents), Multi-Frame Assignment (an approach based on Lagrangian relaxation) and particle filters (a dynamic Monte-Carlo method). Simulators, data readers, metrics, tutorials, documentation etc already exist and are being developed and enhanced by a global team of developers. Liverpool and Dstl (part of UK Ministry of Defence) are at the centre of that community.

Stone Soup is already being used by industrial and government organisations. Stone Soup therefore provides a route to impact for research related to advancing the state-of-the-art in relevant algorithms. Indeed, UoL have approximately £2M of current grants (mostly from Dstl) to develop and apply Stone Soup in contexts related to sonar, astrodynamics, maritime surveillance and video tracking with further funding very likely. A current focus of development for Stone Soup is sensor management, ie real-time optimisation of the control of sensors (eg dynamically steering UAVs and the cameras on such drones to maximise the chance that they detect behaviours of interest). In that context, Stone Soup is relatively sparsely populated with algorithms. While there is funded work to investigate the utility of Monte-carlo tree search and deep reinforcement learning for Stone Soup, this would seem to provide an opportunity for the Department of Computer Science to engage. Similarly, there is a growing interest in developing algorithms that are robust to deception by an intelligent adversary such that game theory is a topic of current interest to the Stone Soup development community.