1 A Novel System for Non-Cooperative UAV Sense-And-Avoid Leopoldo Rodriguez Salazar AEROTECH Systems Ltd Milton Keynes, Buckinghamshire, MK11 1BY, United Kingdom Roberto Sabatini, Subramanian Ramasamy, Alessandro Gardi Department of Aerospace Engineering, Cranfield University Cranfield, Bedfordshire, MK43 0AL, United Kingdom (r.sabatini@cranfield.ac.uk) ABSTRACT This paper presents the research activities performed to develop a Sense-And-Avoid (SAA) system for non- cooperative Unmanned Aerial Vehicle (UAV) collision avoidance tasks. The SAA system utilises Forward Looking Sensors (FLS) and low-cost (and low- weight/volume) Navigation and Guidance System (NGS), composed by a Vision Based Navigation (VBN) sensor integrated with a Global Navigation Satellite System (GNSS) using L1 GPS/GALILEO signals and a MEMS-IMU (Micro-Electromechanical System Inertial Measurement Unit) sensors. The SAA system performs obstacle detection and tracking by combining measures from different FLS types, including both passive and active systems. The process starts with navigation-based image stabilization and considers the use of image morphology operations to extract relevant features that may represent a collision threat. Obstacle tracking is performed in two stages: a low-level tracking stage that utilises an Ad-hoc Viterbi algorithm with four filter branches allowing obstacle classification, in addition to position and heading estimation. High-level tracking is performed with the use of a Kalman Filter that allows an estimation of the velocity of the obstacle by considering the position and heading estimates from the Viterbi algorithm. Sensor fusion is implemented with a Track-To-Track (T 3 ) algorithm using cross-correlation estimation techniques. Risk of Collision (ROC) is assessed with the integration of the joint Probability Density Function (PDF) over a risk area of the UAV and the intruder’s relative positions. The decision-making follows a Boolean logic that considers the output of the ROC estimation and the output of the Ad-hoc Viterbi outputs. Finally, an avoidance algorithm is proposed with a modification of the standard Differential Geometry (DG) algorithm that allows a geometric-based resolution considering a minimum separation distance. Simulation case studies are performed considering the AEROSONDEUAV as the SAA host platform and other civil/military platforms (i.e., A320, B747 and TORNADO-IDS) as the flight obstacles. Results showed that the implementation of DG algorithms with heading command is always sufficient to maintain a safe separation distance from other aircraft with moderate to high flight dynamics, when the SAA process (tracking, decision/prioritisation and avoidance loops) is performed from ranges in excess of 500 metres. KEYWORDS Sense and Avoid, Image Morphology, Ad-Hoc Viterbi Filter, Kalman Filter, Differential Geometry, Kalman Filter, Trajectory Prediction, Risk of Collision and Minimum Separation Distance. I. INTRODUCTION Numerous research efforts are on-going to define standards, procedures and safety thresholds that would make routine access of Unmanned Aerial Systems (UAS) to the commercial airspace a reality. The Sense-And-Avoid (SAA) capability is considered of paramount importance since it would grant the UAS a capability comparable or superior to the human see- and-avoid capability assured by the pilots in manned systems. In general the SAA capability can be defined as the automatic detection of possible conflicts and enactment of an evasive manoeuvre to prevent the collision. According to Lai et al. [1], non-cooperative Collision Detection and Resolution (CD&R) for Unmanned Aerial Vehicles (UAVs) is one of the major challenges that need to be addressed to allow a safe unrestricted access to airspace. Ali Shah [2] emphasizes that the most common approach for an obstacle detection system is the use of multi-sensor platforms, including passive devices such as visible and infrared cameras, and/or active devices such as Laser Radar (LADAR) or Millimetre Wave (MMW) radar. Gauci [3] presented a comparison of different candidate sensor technologies for the SAA problem in aerodrome proximity areas. Computer vision systems are prominent players in the SAA field. High level detection and tracking stages make use of stereo vision techniques and an Extended Kalman Filter (EKF). Sabatini [4] proposed the adoption of a multi- sensor platform for the SAA functionality by using passive and active MMW radar, Forward Looking Infra-