TABLE OF CONTENTS
Transcription
TABLE OF CONTENTS
AUVSI SUAS Team Buzzed, Journal Paper Georgia Institute of Technology Faculty Advisor: David Moroniti Date Submitted: May 28, 2014 This paper describes the undertaking of the Georgia Institute of Technology Team Buzzed in the AUVSI SUAS competition. A systems engineering approach was used to understand the competition requirements, derive capabilities, develop a total system solution, and validate this system through testing. The vehicle is capable of autonomous flight and aerial photography of a desired search area. The team developed a system for autonomous target identification and recognition through custom ground control software. The team has demonstrated through testing that this system can achieve all primary and desired secondary tasks. By combining these individual systems and through intensive testing of the collective system, Georgia Tech believes this entry to be a strong contender in the 2014 AUVSI SUAS competition. TABLE OF CONTENTS 1. SYSTEMS ENGINEERING APPROACH ........................................................................................................................3 1.1 Mission Requirements Analysis ....................................................................................................................................3 1.2. Design Rationale ................................................................................................................................................................4 1.3. Programmatic Risks and Mitigation Methods ........................................................................................................5 1.4. Expected Performance ....................................................................................................................................................7 2. UAS DESIGN .......................................................................................................................................................................8 2.1. Aircraft ..................................................................................................................................................................................8 2.1.1. Propulsion ..................................................................................................................................................................................................................... 8 2.1.2. Planform Sizing........................................................................................................................................................................................................... 9 2.1.3. Drag Analysis ............................................................................................................................................................................................................... 9 2.1.4. Stability and Control ................................................................................................................................................................................................ 9 2.1.5. Manufacturing ........................................................................................................................................................................................................... 10 2.1.6. Modifications ............................................................................................................................................................................................................. 10 2.2. Method of Autonomy..................................................................................................................................................... 11 2.2.1. Data Link ...................................................................................................................................................................................................................... 11 2.2.2. Ground System .......................................................................................................................................................................................................... 12 2.3. Imaging .............................................................................................................................................................................. 15 2.2.1. Camera .......................................................................................................................................................................................................................... 15 2.2.2. Gimbal ........................................................................................................................................................................................................................... 15 2.3.3. Video Transmission ................................................................................................................................................................................................ 16 2.4. Target Recognition and Analysis .............................................................................................................................. 16 2.4.1. Computer Vision....................................................................................................................................................................................................... 17 2.4.2. Localization ................................................................................................................................................................................................................ 19 3. TEST AND EVALUATION RESULTS ......................................................................................................................... 19 3.1. Payload Systems ............................................................................................................................................................. 19 3.2. Guidance ............................................................................................................................................................................ 20 4. SAFETY CONSIDERATIONS AND APPROACH ...................................................................................................... 20 TABLE OF FIGURES Figure 1 – SUAS task requirement decomposition and system mapping .............................................................................. 4 Figure 2 – Gantt chart showing planned execution of tasks. The integration milestone between imaging and the airborne system is marked by “Buzz”, Georgia Tech’s mascot. ......................................................................... 6 Figure 3 – The Georgia Tech platform for SUAS, Buzzed ............................................................................................................... 8 Figure 4 – Buzzed model in AVL ............................................................................................................................................................... 9 Figure 5 – Air-Drop release mechanism .............................................................................................................................................11 Figure 6 – UAS ground and onboard systems data flow ..............................................................................................................12 Figure 7 – Piccolo Command Center user interface .......................................................................................................................13 Figure 8 – Air-Drop control GUI..............................................................................................................................................................14 Figure 9 – The team’s custom gimbal with the Sony FCB-EV7500 .........................................................................................16 Figure 10 – Timeline of functions in the target recognition and analysis process ..........................................................16 Figure 11 – Image filtering by principal colors ................................................................................................................................17 Figure 12 – Shape classification by binary edge detection and corner count ....................................................................18 Figure 13 – Feature detection (top), matching (bottom left), and filtering (bottom right) .........................................18 Figure 14 – Before and after of image color simplification ........................................................................................................19 Figure 15 – Flight control station and Piccolo command station diagram ..........................................................................20 TABLE OF TABLES Table 1 – Primary task failure-risk areas and mitigation methods .......................................................................................... 5 Table 2 – SUAS task completion scheduling risk assessment, adapted from MIL-STD-882C ....................................... 7 Table 3 – Expected performance at SUAS ............................................................................................................................................ 7 Table 4 - Parasitic drag breakdown ....................................................................................................................................................... 9 Table 5 – Stability and control properties of Buzzed.....................................................................................................................10 2 Team Buzzed 1. SYSTEMS ENGINEERING APPROACH 1.1 Mission Requirements Analysis The rules specify two primary and eight secondary tasks for the UAS to perform. Accomplishing all ten tasks scores the most points and maximizes success at the fly-off. However, operating under the principle of resource scarcity, it is vital to prioritize tasks and appropriate resources accordingly, including time, labor and funding. The prioritization process was particularly important to the Georgia Tech team as a first-time entrant in SUAS. To determine which tasks to target, each one was decomposed into a set of requirements for the system to perform based on the rules. In absence of competition experience, three criteria were used to evaluate the results of the requirements analysis: 1. The amount of overlap in requirements between tasks 2. The number of requirements added per task 3. The complexity of any added requirements Complexity was rated qualitatively based on the team’s research of past SUAS entries and existing operational UAS. The tasks to perform were prioritized qualitatively based on a combination of the three criteria above and an assessment of the team’s capability and background. The requirements analysis method described above resulted in the framework seen in Figure 1. The team first decomposed the Primary autonomy and search tasks into their requirements. Moving into Secondary tasks is not allowed by the rules until threshold Primary requirements are met, committing the team to all requirements except autonomous landing, regardless of their difficulty level. The team then derived a high-level architecture that can fulfill the system requirements, resulting in an autonomous UAS with imaging and computer vision (CV) capabilities, and a ground station to provide a man-to-system interface. With an architecture chosen to fulfill the Primary tasks, the Secondary tasks were decomposed to determine additional derived requirements, their complexity, and how they map to the high-level architecture, as seen in Figure 1. Many of the Secondary tasks overlap with the Primary requirements, but only added requirements are graphically illustrated in Figure 1 for simplicity. Dashed lines from tasks to requirements denote an optional addition. Tasks that only required a single, relatively simple addition were prioritized for completion. More complex tasks were targeted as optional, with their completion dependent on progress of simpler tasks. The overall result of the mission requirements analysis was a tiered approach, with a progression in task number and complexity as the system is developed. This approach enabled the team to focus its efforts in a structured manner towards the completion of as many tasks as possible while taking into account its limited experience, resources, and time. 3 Team Buzzed Figure 1 – SUAS task requirement decomposition and system mapping 1.2. Design Rationale The mission requirement analysis seen above showed that SUAS is substantially focused on system integration. The selected architecture is a platform to complete as many functions as possible; there is no preference for a specific shape or features as long as the architecture lends itself well to integrating many sub-systems. This is exemplified by the variety of craft fielded at SUAS, including fixed-wing airplanes, helicopters, and multi-rotors of all sizes. This functionallyoriented view led to the team’s goal of quickly fielding a working platform, spending more time on task-related sub-system development and test than vehicle- or software-related design. This rationale lends itself strongly to the integration of off-the-shelf systems. As a first-time entrant in SUAS, the team also sought to simplify wherever possible. This led to choosing operational, “known- 4 Team Buzzed good”, and verified systems, including some commercial off-the-shelf (COTS) components for major elements of the high-level architecture: 1. Aircraft: an in-house built gas-powered medium-endurance aircraft, with many hard-points and large payload capacity, which has already been flight-proven as a stable, reliable platform for the airborne systems. 2. Autopilot: Cloud Cap Piccolo, a professional product with near plug-and-play capability serves as an integrated solution for nearly all autonomy tasks required by the competition. 3. Ground System: a combination of in-house built stations that have been flight-tested with the chosen aircraft, and COTS stations to support the autopilot, camera, and gimbal, all operated by trained personnel. 4. Computed Vision: a MATLAB-developed toolbox for shape and character recognition with flexible format inputs and detailed online tutorials and support. 1.3. Programmatic Risks and Mitigation Methods The programmatic risks in SUAS can be classified into schedule-related and failure-related risks. Scheduling risks include overruns in the critical path and underestimating labor times, manufacturing times, or shipping lead-times. Failure-related risks include hardware failures that result in damage to the aircraft, camera, or instrumentation, and task-completion failures such as being unable to classify a target. Any failures also tie back to the schedule and can cause significant overruns, but their occurrence cannot be predicted and worked into the schedule a priori. The approach is thus left to identify potential high-risk areas and have a detailed plan to mitigate them. The major failure-risk identification and mitigation for Primary tasks is detailed in Table 1. Although Secondary tasks also have risk, the Primary tasks pose the greatest programmatic risk because failure in those areas precludes the team from attempting Secondary tasks. Failure-risk mitigation via safety protocols is detailed in Section 5 – Table 1 is for programmatic risk mitigation. Table 1 – Primary task failure-risk areas and mitigation methods Failure-Risk Areas Loss of aircraft function Loss of autopilot function Loss of gimbal/camera command and control First-time aircraft-autopilot integration Inability to classify >2 features per target Mitigation Method Use a verified, stable, in-house built airframe with over 45 successful flights. Use a small UAS industry-leading autopilot system with built-in failsafe features. Pass gimbal control signal through the autopilot system; use a reliable camera. Perform hardware-in-the-loop (HIL) testing, then use a small, inexpensive aircraft to test autopilot functions and reduce the consequence of failure. Use a commercially available, well-supported CV code with demonstrated similar use-cases. The failure-risk mitigation strategy echoes the design rationale detailed in Section 1.2: using known-good operational systems or COTS components is a risk reduction method and reduces the amount of uncertainty in the system performing reliably. The same rationale also deliberately eliminated much of the design and development for the major systems, freeing time for sub-system development and test and reducing scheduling risks substantially. As a consequence, the Gantt chart shown in Figure 2 does not include any aircraft design and manufacture or major 5 Team Buzzed programming efforts for the autopilot and CV systems. Instead, the chart only focuses on completing the task-related requirements. Figure 2 – Gantt chart showing planned execution of tasks. The integration milestone between imaging and the airborne system is marked by “Buzz”, Georgia Tech’s mascot. The Gantt chart follows the flow dictated by the mission requirement analysis in Section 1.1, which prioritized Primary, then simple Secondary, and finally complex Secondary tasks. Figure 2 shows two scheduling critical paths, which map to the high level architecture of the aircraftautopilot and camera-CV systems. Any underestimation in any phase of the schedule trickles delays to tasks along the critical paths. This means that the tasks found later in the critical path become more susceptible to schedule delays. However, the consequence of delays becomes smaller for increasingly complex tasks because of how they were prioritized: the goal for the team as a firstyear entrant is to complete as many tasks as possible, leaving more complex tasks for future years if necessary. This risk management method is graphically displayed in the risk assessment matrix seen in Table 2, where tasks of increasing complexity flow from top-left to bottom-right. This means that the critical Primary tasks are the most likely to be completed, while less-critical Secondary 6 Team Buzzed tasks are more prone to scheduling delay. In all, no tasks were at an imperative level, meaning the risk management strategy is well-suited for the team’s objectives. Table 2 – SUAS task completion scheduling risk assessment, adapted from MIL-STD-882C Susceptibility to Scheduling Delay Severity of Consequence Impossible Improbable Remote Critical 7.1 7.2 Marginal 7.6, 7.9 7.8, 7.5 Occasional Probable Frequent 7.4 Acceptable 7.10 Negligible 7.3 7.7 Risk Code / Action - Imperative to suppress to lower risk level - Take action to mitigate, while balancing design goals - Operation permissible 1.4. Expected Performance The team’s tiered approach to targeting tasks has placed substantial effort on achieving the Primary tasks. At the time of this writing, that effort has translated to progress in both the autonomous flight and computer vision areas, giving the team a high degree of confidence of meeting at least the threshold requirements, if not the objectives of the Primary tasks. This probability of attempting the ten different tasks is listed in Table 3. Table 3 – Expected performance at SUAS Task # Primary Secondary Task Name Probability of Attempt 7.1 Autonomous Flight High 7.2 7.3 Search Area Automatic Detection High 7.4 Actionable Intelligence 7.5 Off-Axis Target High 7.6 Emergent Target High 7.7 Remote Information Center 7.8 7.9 Interoperability Infrared Search 7.10 Air-Drop Low Medium Low High High Medium Many of the Secondary tasks overlap with the Primary ones and only have small, manageable additions, increasing the team’s confidence in completing tasks 5, 6, 8, and 9. The Actionable Intelligence (classifying all features) task depends on the progress of the Primary task, but enough 7 Team Buzzed information is expected to be collected for a complete analysis if scheduling allows sufficient testing time. The Air-Drop requires more hardware additions than most Secondary tasks, but the aircraft used for the competition was originally designed to drop a payload on a target. With some modifications, it is likely the team will attempt the task, even if logic for an autonomous drop is not added. Finally, the Automatic Detection and Remote Information Center tasks call for a more complex development, making them a lower priority for the team at SUAS. 2. UAS DESIGN 2.1. Aircraft Buzzed is a blended wing aircraft using an H-tail configuration, tricycle landing gear, and a gas-powered engine arranged in a tractor configuration. This design, pictured in Figure 3, has been used by the Georgia Tech Design Build Fly team in two separate competitions completing 45 total flights, proving to be a reliable design. Due to the nature of the past two competitions, Buzzed was designed for ample stability under conditions of high wing loading and an off-centerline center of gravity. Stability and reliability are keys to enabling more focus to be put forth on the subsystems by minimizing risk factors caused by the aircraft, making Buzzed an ideal candidate for the 2014 AUVSI-SUAS competition. Figure 3 – The Georgia Tech platform for SUAS, Buzzed 2.1.1. Propulsion The fully utilize the 40-minute flight time allotted in the demonstration period, a 0.46 cubic inch two-stroke gas engine was selected with a large enough fuel tank. Gas was chosen over electric components because of its high energy density, reducing weight for at an equivalent flight endurance. To optimize the performance of this motor multiple propellers were mounted to the engine on a static thrust stand to determine their thrust. The ideal propeller is capable of providing enough thrust for takeoff while operating at low RPM during cruise for better propulsion system efficiency. The results of testing indicated that an 11.5 x 6 propeller will be the most efficient propeller for the selected engine. \ 8 Team Buzzed 2.1.2. Planform Sizing The blended-wing body design is more efficient than a fuselage and provides flexibility in hardpoint attachments. The large wing area of 11 ft2 enables cruise at low speed, useful in reducing motion blur for image capturing. Athena Vortex Lattice (AVL) was used to size the wing, empennage and control surfaces. AVL is a code developed at MIT that calculates the aerodynamic characteristics of an airplane by discretizing the wing into a vortex sheet along the span and camber lines and applying boundary conditions at the wingtips and trailing edges. The Buzzed configuration in AVL is seen in Figure 4. This virtual model was used to estimate stability and control characteristics, which were also needed for initial gain tuning and flight simulation with the autopilot. Figure 4 – Buzzed model in AVL 2.1.3. Drag Analysis A parasitic drag estimate was computed by summing each component’s drag contributions, approximated using empirical estimation techniques in Hoerner’s Fluid Dynamic Drag, and then normalizing each component according to the wing reference area. Figure 4 shows the contributions of the main aircraft components. The induced drag was estimated from AVL. The blended wing design provides high efficiency and low drag to reduce fuel consumption during the long-endurance mission. Table 4 - Parasitic drag breakdown Part Drag Percent of Total Wing 0.0147 64 Landing Gear 0.0050 22 Horizontal Tail 0.0025 11 Vertical Tail 0.0008 3 Total 0.023 100% 2.1.4. Stability and Control To ensure that the aircraft can successfully complete the design mission, both static and dynamic stability characteristics were computed in AVL. This information was combined with the principal moments of inertia found in CAD to determine dynamic stability behavior using the full 6 DOF linearized, coupled differential equations found in Philips Mechanics of Flight. The most 9 Team Buzzed important static derivatives, deflections, and the static margin are seen on the left side of Table 5, while the most important high-frequency dynamic mode behavior is seen on the right. The static evaluation confirms that the aircraft is statically stable with 9.6% margin. The dynamic analysis indicated that the aircraft is stable in all high-frequency modes, with damping ratios and frequencies within the expected ranges for small unmanned vehicles. This stable platform lends itself well to being stabilized and controlled by an autopilot. Table 5 – Stability and control properties of Buzzed Static Stability Wtotal(lbs) Inputs V(ft/s) CL Aerodynamic α (deg) Parameters β (deg) Cm,α (rad-1) Stiffness Cl,β (rad-1) Coefficients Cn,β (rad-1) Static Margin % Chord 25 50 1.0 7.7 0.0 -0.446 -0.134 0.041 9.6 Dynamic Stability Mode Short-Period Dutch Roll -1 Damping Rate (s ) 2.297 0.324 Time to Half (s) 0.302 2.143 Damping Ratio 0.591 0.110 Damped Freq. (s-1) 3.137 2.934 -1 Undamped Freq. (s ) 3.888 2.957 Control -1 Cl,δa (deg ) 0.035 δa (deg) Cm,δe (deg-1) -0.001 δe (deg) Roll 2.777 0.250 0 -4.1 2.1.5. Manufacturing To decrease vibration from the motor and get a more stable image, the team used a Hyde Motor Mount. Data provided by the manufacturer claims that the Hyde mount lowers the vibrational amplitude by 70%. Experimental results indicated that a 1/8 inch poplar plywood nose box was sufficient to withstand the torque and static thrust of the motor. This box was also used as housing for the fuel tank and various electronics such as the servo motors controlling throttle and the nose gear. The nose gear itself is also attached to the engine mount. The main landing gear was fabricated from a solid piece of aluminum attached to the fuselage via screws. Landing gear placement was dictated by CG position. The empennage is attached using two carbon fiber tubes attached to the carbon fiber wing spar. The H-tail has an elevator and a dual rudder system controlled by pushrods actuated by servos. The control surfaces are attached using a socket style hinge to minimize drag. A laser cutter was used to print the aircraft from sheets of balsa wood and plywood, ensuring accuracy between the CAD designs and the final aircraft. This is especially critical for ribs, as twist introduced in construction can make the airplane difficult to control. The ribs were specifically designed to fit together accurately like a jigsaw puzzle, allowing for repeatable and accurate construction. 2.1.6. Modifications As it stood, the Buzzed platform required several modifications to make it compatible for the 2014 AUVSI-SUAS competition. A larger fuel tank was fitted to the aircraft to allow 40 minutes of flight, eliminating the need to land for refueling. The camera gimbal detailed in Section 2.2.2 was designed to attach to a pre-existing hard-point on the bottom of the aircraft, substantially reducing the number of modifications needed. The wing was strengthened around the mount for the pitot 10 Team Buzzed tube, which is required by the Piccolo autopilot system to calculate airspeed. The aircraft had sufficient internal space for the remaining Piccolo components. A drop mechanism for a different competition was originally mounted at the gimbal hardpoint. To attempt the Air-Drop mission, the drop was relocated to the wing. The laser cut claw seen in Figure 5 was chosen to secure and release the payload. The claw is composed of two laser cut arms whose base act as a gear and are controlled by a small servo. The claw opens to 2.5 inches and can carry the required payload size and weight. Using simple components minimized weight and moving parts in the wing. Figure 5 – Air-Drop release mechanism 2.2. Method of Autonomy To be able to accomplish the autonomous flight primary task objectives, as well as to accomplish several of the secondary task objectives, an autopilot flight system is required. The autopilot system must be capable of autonomous take-off and landing, flight, waypoint tracking, and respond to flight plan changes during flight while staying within a specified flying zone. The autopilot system must also be reliable, as a failure in this system can cause a catastrophic crash. To fulfill these requirements a Piccolo SL Autopilot System has been chosen. In addition to meeting task requirements, members of Buzzed are familiar with the Piccolo software, reducing the time required to incorporate the software into the aircraft. The Piccolo SL was ideal for the competition as the autopilot system is small, lightweight and has all the capabilities necessary to successfully complete the mission tasks. It uses a combination of GPS antenna, Inertial Measurement Unit, pitot-tube pressure sensors to determine position, altitude, orientation, attitude, and airspeed. These inputs allow the autopilot system to fly the aircraft autonomously. 2.2.1. Data Link The many data links of the UAS are separated throughout the spectrum so that they do not interfere with each other. In the unlikely event of one of these data links failing, the team has established safety procedures to debug and re-establish connections. 2.2.1.1. Autopilot The Main Data Link is a 900MHz connection that is to be used for communications between the Piccolo Portable Ground Station (PGS) and the Piccolo SL Autopilot System. 11 Team Buzzed This link is the main channel of communication between the Flight Control Station and the Piccolo SL Autopilot System. The safety pilot transmitter and gimbal operations station also connect with the Piccolo SL Autopilot System through this link for manual control of the aircraft and gimbal respectively. 2.2.1.2. First-Person View (FPV) Camera A forward-facing FPV camera is used as a safety factor in case of possible component failure. If the autopilot system fails while the aircraft is out of sight, the forward facing camera can be used to fly the plane manually. The video taken from the first person view camera is transferred over an Immersion RC 5.8 GHz Video Transmitter. This transmitter was chosen for its ability to transmit video over a wide range of frequencies with a power output of 600 mW and very little noise. 2.2.2. Ground System The Ground Station will be composed of the Piccolo Portable Ground Station (PGS) provided by Cloud Cap with the Piccolo SL system, along with several computer stations running dedicated software based on the task given to the operator. Figure 6 – UAS ground and onboard systems data flow 2.2.2.1. Flight Control System (FCS) The operator of the FCS is tasked with monitoring and providing commands to the Piccolo SL Autopilot System by interfacing with the Piccolo Command Center (PCC) software installed on the Flight Control Computer (FCC). Through the PCC, the operator can monitor altitude, airspeed, GPS position, heading and attitude while being able to command the autopilot, set flight boundaries and limits, and provide or modify a flight plan during 12 Team Buzzed flight. These are all important tasks as they will allow the team to fulfill primary and secondary mission objectives. The FCC is linked to the Piccolo PGS through a serial connection which in turn links to the Piccolo SL Autopilot through the Main Data Link. The lost communications waypoint safety feature is also set by the FCS operator through the PCC. If the FCC crashes, the operator will have a backup FCC running in parallel so they can simply plug the backup to the Piccolo PGS, download the most recent data from the Piccolo SL Autopilot and resume operations while the main FCC is inspected and rebooted. During each flight, telemetry data is automatically saved and can be accessed at a later time. Figure 7 – Piccolo Command Center user interface 2.2.2.2. Safety Pilot Station (SPS) The safety pilot is tasked with manually controlling the aircraft in any situation or circumstances where the autopilot cannot. The SPS is composed of the First Person View (FPV) computer and the safety pilot transmitter. The FPV computer is directly linked to the FPV system on board the aircraft through the FPV Video Link and is a safety feature added in case the aircraft flies out of the line of sight of the safety pilot and they need to take manual control of the aircraft. Manual control of the aircraft can be attained by the safety pilot through a switch on his transmitter. The transmitter is directly linked to the Piccolo PGS and communicates with the Piccolo SL through the Main Data Link. 2.2.2.3. Payload Operations System (POS) The GOS operator is in charge of the gimbal control and monitoring software. The GOS is composed of Gimbal Control Computer (GCC) computer and the Gimbal remote control (GRC). The GCC receives a live feed from the Sony block camera on board the 13 Team Buzzed aircraft via the Gimbal Video Link. The GRC controls the gimbal movement through software installed on the GCC. The GCC is connected to the payload pass-through port on the Piccolo PGS and controls the gimbal through the Piccolo SL Autopilot. 2.2.2.4. Image Recognition Station (IRS) The IRS operators are tasked with maintaining and monitoring the IRS. The IRS is composed of several computers running image recognition software written by the team in parallel to find and identify all targets characteristics autonomously. 2.2.2.6. Air-Drop Control (ADC) During estimation of the payload drop location; position and velocity of the aircraft are taken into consideration. External factors such as wind, the drag profile of the payload during the drop trajectory, and the data rate that informs the ground-station of the state of the aircraft are also taken into consideration. The first objective of the mission model is to provide an estimation of the drop location. Using basic physics the payload follows: assuming no external interference. From the estimated altitude a computer model determines the time, t, that is required for the payload to reach the ground. Using this calculated time, the model then calculates the distance traveled along the ground. This portion of the model introduces a constant, k that is multiplied to the distance traveled along the surface in order to attempt to model the effects of drag on the payload during the drop trajectory. The ground station uses the data obtained from the on-board sensors to create a GUI that indicates the current state of the aircraft and the desired drop location of the payload. The desired drop location of the payload, recorded before the start of the mission, is indicated by the red ‘X’. The blue ‘X’ indicates the estimated location of the aircraft. A black line connects the last estimated location of the aircraft with the desired drop location. A black ‘X’ indicates the predicted drop location. Figure 8 – Air-Drop control GUI 14 Team Buzzed 2.3. Imaging 2.2.1. Camera For this competition a camera must have a pixel density or zoom large enough to identify a letter on a sign and reduce the effects of plane vibration as much as possible while being able to transmit an image quickly. The system must also be able to identify a heated sign in order to successfully complete the mission. High definition video cameras and Digital Single-Lens Reflex (DSLR) cameras capable of both high pixel density and high magnification zoom are the two most commonly used devices in unmanned aerial reconnaissance. DSLR cameras typically provide a very high-resolution image, a large range of manual zoom capabilities, high shutter speeds, and a large array of manual settings. Video cameras provide a real time, high definition stream with moderate zoom capabilities. Both camera types are capable of image stabilization, which is critical for an aerial camera. The main differences between the two types are that a DSLR is most efficient for still images while an HD camera constantly streams video, sending constant data to the operator. This difference enables the HD camera to stream nearly in real-time, allowing a faster reaction to the aircraft’s environment. HD cameras also have a smaller form-factor and weigh less than DSLRs, making them easier to accommodate on the aircraft. The real-time and form-factor advantages led the team to select HD cameras over DSLRs. A Sony FCB-EV7500 HD video camera was chosen due to its small size, low weight, high quality imaging capabilities, and built in software features. The camera allows for near real-time 1920x1080p video streaming, 30x optical zoom, and advanced image stabilization, as well as a built in infrared capabilities for identifying heated objects. These capabilities provide the team with high quality imaging for the Search Area, Infrared, Emergent Target, and Actionable Intelligence tasks. Ground tests have indicated very clear images even at a worst-case condition of full zoom with external vibration present. 2.2.2. Gimbal The gimbal is designed to provide 180 and 360 degrees of pitch and yaw freedom respectively, creating a complete hemisphere of target tracking under the aircraft. Pitch and yaw control is necessary design because after identifying a target the Sony FCB-EV7500 will be zoomed in for increased target recognition abilities, and will likely need to be reoriented to the target as the aircraft moves. Increased control provides the team with both a method of scanning the searchable area without having to execute multiple flybys and zooming on the target to increase the ease with which it is identified. The main constraints on the gimbal are to ensure it does not interfere with takeoff and landing, and being of minimal size and weight to reduce its negative effects on aircraft flight. Two Futaba S3003 servos were selected to control the pitch and yaw of the gimbal based on their torque and angular resolution, while a slip ring provides the gimbal with infinite rotation around the yaw axis without tangling wires. To satisfy the gimbal constraints and securely fix the servos and camera into place, the gimbal was custom designed and built utilizing precision laser cutting and additive manufacturing capabilities. The gimbal was 3D printed in three separate sections, and assembled to include steel ball bearings to create a mechanical slew ring. The gimbal achieves yaw rotation via a servo that is geared with two laser-cut acrylic gears. The complete gimbal system can be seen in Figure 9. The gimbal camera’s pointing direction can be interpolated from the commanded position of the pitch servo, and from an encoder located on the upper surface of the gimbal, providing yaw information. 15 Team Buzzed Figure 9 – The team’s custom gimbal with the Sony FCB-EV7500 2.3.3. Video Transmission Target acquisition video from the gimbaled camera is sent to the ground station using a Microhard 2.4 GHz Wi-Fi radio. This model was chosen for it compact, lightweight design. It weighs 24 grams with 1 watt of RF output and up to 12 mbps of bandwidth. This model’s major feature is its long-range capabilities, with a line of sight range up to 14 miles. Using Wi-Fi radio to transmit video to the ground station is the fastest and most reliable way to send large packets of data over large distances and will limit the interference caused by other teams. In order to stream the video to the ground station with minimized lag the system uses an Airborne Innovations h.264 video encoder board, which compresses 1080p video at up to 30 frames per second. 2.4. Target Recognition and Analysis The recognition of the target and its analysis is a complex process that involves hardware, software and human interaction. Figure 10 below displays the flow of the recognition and analysis process with respect to time. The on-board hardware and its functions were described in Section 2.3, while the analysis functions are described in detail in the sections to follow. Figure 10 – Timeline of functions in the target recognition and analysis process Target recognition begins with the gimbal system operator captures an image of the target for classification analysis. When the image is captured, the onboard systems record the aircraft GPS position, altitude above ground level (AGL), heading angle, gimbal angles, and aircraft motion angles for localization and orientation analysis. After these values are obtained they are wirelessly transmitted along with the captured image to the ground computational unit (GCU). The GCU consists of a directional receiver antenna to support long-range Wi-Fi 16 Team Buzzed connections, and a portable computer with MATLAB code that performs all desired image recognition and analyses. The results from the CV analysis are compared to a human’s visual analysis of the image to confirm or deny the automated output, ensuring accurate recognition in cases of false computer detections. 2.4.1. Computer Vision The Computer Vision (CV) algorithm is a combination of functions that perform a detailed analysis of the input images to output the four visually-based features required by the for the Search Area, Actionable Intelligence, and Infrared tasks: 1. 2. 3. 4. Alphanumeric character Shape of platform backgrounds Character color Background color Images acquired by the gimbal operator are stored in a folder on the GCU. The CV code then automatically reads every new file and performs the analyses detailed below. 2.4.1.1. Image Filtering Before any shape or letter recognition, the image is filtered to reduce background noise. Good filtering substantially increases the success rate of the required recognition tasks. The filtering function segments the input image into multiple images based on color strength. Color distribution data are obtained from the image and used to split the main colors into separate images. An example is shown in Figure 11 below. A baseline image is split into its 6 strongest principal colors to give cleaner images, which are passed through to the next phases of analysis. Figure 11 – Image filtering by principal colors 2.4.1.2. Shape Recognition The filtered images are converted into binary (black and white) images, to which an edge detecting function is applied. The edge detection is performed through neighboring pixel comparison, where steep local changes in pixel values are classified as an edge. These edges can create outlines of geometric shapes with a distinct amount of corners. If the edges form an enclosed space within angular and Cartesian limits, the algorithm counts the number of corners by finding discontinuities in a spline fit of the edges. The corners which define the target shape are used to crop the original image to a manageable size for color and letter recognition as well as better corner recognition. The number of corners is used to classify shapes, since common polygons have a distinct number of corners. An example of the shape detection process is shown in Figure 12 below. 17 Team Buzzed Figure 12 – Shape classification by binary edge detection and corner count 2.4.1.3. Letter Recognition A copy of the cropped image acquired from the shape recognition process is sent to the letter recognition code. The target image will rarely be a direct overhead image, so it is straightened for easier analysis. The angle of deflection between vectors of straight lines that are nearly horizontal and horizontal reference of the image is computed and used to rotate said vectors to the horizontal. To recognize the letter, the filtered and rotated images are compared to the alphabet templates one-by-one until there is a match as seen in Figure 13. The comparison is performed using feature detection which extracts feature components of the template letter, calculates their centers (seen as the green circles in Figure 13), then groups these feature components by their coordinates. Figure 13 – Feature detection (top), matching (bottom left), and filtering (bottom right) The last step in letter recognition process is to perform a feature match. The comparison results in an output file that contains the number of identical points. Some of the results might be inconsistent or erroneous, so a filter which erases points that do not follow an overall pattern, is applied. The template that had the most feature matches with the image is chosen as the letter. 2.4.1.4. Color Recognition A copy of the cropped image acquired from the shape recognition process is sent to the color recognition code. The cropped target image is initially read as Red-Green-Blue 18 Team Buzzed (RGB) space, is simplified and transformed into Hue-Saturation-Value(HSV) space, then compared to a color-map and classified under a major color hue for both the shape and letter. In order to differentiate between the background color, the shape color and the letter color, the algorithm uses the first pixel in the image as a reference, and is compared to all other pixels in the image. If this background pixel is different from the new pixel selected, and the pixel is for the first time, then the pixel color is saved as the shape color. If the new pixel color is different from both the shape pixel color and the background pixel color, then the new pixel color is saved as the letter color, the algorithm ends, and exports the shape color and the letter color to an excel file. Figure 14 – Before and after of image color simplification 2.4.2. Localization 2.4.2.1. Position In order to find the GPS coordinates of a ground target, the aircraft’s direction and position must be obtained from the Piccolo software while the gimbal direction must be interpreted from the servo commands. Using the orientation of the aircraft, as well as the direction the camera is pointing, the relative direction of the image can be calculated. This relative direction, in conjunction with the altitude of the aircraft, can be used to find a ground level distance to the target from the aircraft. Combining the ground distance between the aircraft and the target with the GPS coordinates of the aircraft yield the position of the target. 2.4.2.2. Orientation To perform recognition of image orientation two data sets are required: a magnetometer reading at the time when image was taken, and the angle of rotation for the image calculated in the section 2.4.1.3. The magnetometer reading determines deflection of the heading vector of the plane from the magnetic north. The image rotation angle gives the deflection from the plane’s heading vector to the target’s baseline. Thus the sum of these deflections provides the total vector angle from the cardinal north direction to the heading of the target image. The result must be rounded to the nearest 45 degrees and compared to each cardinal direction. The cardinal direction is acquired from the division of the angles into 45 degree segments. When the result is the same angle as the represented cardinal direction, the direction is output. 3. TEST AND EVALUATION RESULTS 3.1. Payload Systems Based on the results from extensive testing of the image recognition software and the gimbal system operations as well as the past experience of the Buzzed platform all primary objectives and desired secondary objectives will likely be met. 19 Team Buzzed Image recognition has provided promising results by successfully identifying letters, shapes, and colors on different backgrounds. The threshold requirements for the search area tasks have been met and the next phase of development will be to incorporate the gimbal angles, computer vision, and aircraft orientation to derive the sign locations. The selected camera has also successfully demonstrated its ability to locate a heated target using its infrared capabilities. Based on the camera’s tested imaging capabilities and the gimbal’s controllability, the emergent target is likely to be found. 3.2. Guidance The process to achieve autonomous flight was an extensive one. First, an Athena Vortex Lattice (AVL) model of the aircraft had to be developed and a software simulation done. The software simulation and AVL model served to establish a baseline for the Autopilot control gains and coefficient values. The next step is to perform a hardware simulation. The Piccolo SL Autopilot System and avionics needed to be integrated to an aircraft to perform a hardware simulation. The Skyhunter FPV UAV Platform was used for preliminary testing and training as the aircraft has a similar amount of control surfaces to our competition aircraft and its frame would protect the avionics in case of an emergency. The hardware simulation served to validate the baseline and to further tune the system before a test flight was performed. Finally, a test flight was performed in a safe area, closely following a pre-established flight plan to test out the Piccolo SL Autopilot System. Figure 15 – Flight control station and Piccolo command station diagram The same process was followed to integrate the Piccolo SL Autopilot System into the competition aircraft. Simple maneuvers like were performed during the initial test flights while tuning the control gains, subsequent test flights involved more complex maneuvers for example autonomous take-off and a closed search ladder circuit within a set boundary limit that involved waypoints at different altitudes. 4. SAFETY CONSIDERATIONS AND APPROACH A successful safety plan must be thorough, consistent, and practiced extensively without deviation. Buzzed’s safety plan includes elaborate pre-flight and post-flight checklists, battery monitoring, and crew familiarity with the aircraft. The flight crew is extremely experienced with operating the aircraft having flown over 45 flights following the same strict safety plan with each 20 Team Buzzed flight. The batteries are brightly colored for quick identification in the unlikely event of a crash. A system is also in place to keep track of battery life by monitoring charges and the general health of the batteries. Buzzed has multiple redundant systems in place to minimize the likelihood of a single failure to cause mission deviation. Although AUVSI-SAUS competition flight is autonomous, a safety pilot has the ability to override the autopilot system at any time and the authority to do so if he or another crewmember finds the need to do so. Risk Probability of Occurrence Consequence Ground Station Piccolo Software Crash Moderate - Software occasionally crashes. Low - Reboot Piccolo software Automatically reconnects to the aircraft. Computer Crash Improbably - Computers show no issues with software Low - Backup computers present which ru essential flight software in case of failure. Primary Comm loss > 10 sec Improbable - Comms tested at competition ranges Moderate - Vehicle shall automatically return home. Primary Comm Loss > 3 min Improbable - Comms tested at competition ranges Moderate - Terminate flight via an autonomous landing. Loss of Payload Comms Improbable - Comms tested at competition ranges Moderate - Check ground station antenna angle, restart modems and receiver. Power Failure at Tent Improbable Low - All computers and ground station components have two hours battery life. Safety Pilot Loses sight of Aircraft Moderate – Distance and orientation can reduce visibility Low - Pilot utilizes FPV system to return to field - If FPV fails, pilot commands autonomous return to home. Aircraft Power failure on piccolo Improbable - Battery life condition checked every charge Severe - Aircraft systems failure. Motor failure Improbable - Motor tested and used often before flight. Severe - Attempt manual unpowered landing. Camera failure Improbable - Camera were extensively tested before flight. Moderate - Reboot camera software - If unable use FPV to return home. Gimbal failure Improbable - Structure and servos were extensively tested before flight. Moderate - Use FPV to return home. Loss of Control Surface Improbable - Preflight and postflight checks to vital aircraft components. Moderate - Redundant control surfaces. Subsystem battery failure Improbable - Battery life has been meticulously tracked during normal usage. Severe - Terminate Flight via Autonomous Landing. 21 Team Buzzed