Marcellin, Michael W.; Hung, David; Tang, Cinthya; Allred, Coby; McKeever, Kennon; Murphy, James; Herriman, Ricky; Univ Arizona, Dept Elect & Comp Engn (International Foundation for Telemetering, 2017-10)
      The Arizona Autonomous Club is a student organization at the University of Arizona which designs, builds, and competes with Unmanned Air Systems (UAS). This year, a 25% scale Xtreme Decathlon model aircraft was selected and successfully converted into a fully autonomous UAS for the AUVSI Student Unmanned Aerial Systems (SUAS) 2017 competition. The UAS utilizes a Pixhawk autopilot unit, which is an independent, open-hardware project aiming at providing high-end autopilot hardware at low costs and high availability. The Pixhawk runs an efficient real time operating system (RTOS) and includes sensors such as a GPS unit, IMUs, airspeed, etc. The UAS also includes an onboard imaging system, which is controlled by an onboard computer (OBC). The Pixhawk and OBC are interconnected with two ground control stations (GCS) using the Robot Operating System (ROS) framework, which is capable of extending overall system capabilities to include an expanded telemetry downlink, obstacle avoidance, and manual overrides.

      Marcellin, Michael W.; Hung, David; McKeever, Kennon; Ramirez, Ricardo; Univ Arizona, Dept Elect & Comp Engn (International Foundation for Telemetering, 2017-10)
      Accurate image classification is one of the core challenges in computer vision. At the annual AUVSI SUAS competition, this challenge is faced in the form of ground target classification from an unmanned aerial vehicle (UAV). Additionally, due to the constraints imposed by the UAV platform, the system design must consider factors such as size, weight, and power consumption. To meet performance requirements while respecting such limitations, the system was broken into two subsystems: an onboard subsystem and a ground based subsystem. This design allows the onboard subsystem, comprised of a DSLR camera and single-board computer, to capture ground target images and perform rudimentary target detection and localization. For further processing and to ultimately classify the targets in each image, data packets are sent to the ground-based subsystem via a 5 GHz wireless link. Convolutional networks are utilized on the ground to achieve state-of-the-art accuracy in classification.