2001Flight Trials of the Continuous Visual Navigation System

2001Flight Trials of the Continuous Visual Navigation Systemreasons. on flat terrain where tcn is ineffective andDefence, through the Applied Research Programme. Theground-based GPS jammers are most effective!)goal was a linear feature matching technique, whichare usable even when flying very low (an importantmilitary requirement)s not reliant upon linear feature intersectionscan be detected by simple and well established imageequiresngprocessing techniques (although complex techniquesare needed and databases are quick and easy tocan produce better resultsproduce)are easily extracted from satellite and photographicimposes a relatively low computational load and ismagery (and consequently database production isleap to implementquick and inexpensive and can incorporate valuableis robust to uncertainty in fix position (i.e. can copetarget information). Figure 1 shows a screen grab fromwith ambiguous linear features)a software package which has been specificallydesigned to make the process of generating a CVNAs for all linear feature matching systems, an operationaldatabase from any satellite imagery a simple and quickCVN system requires the following componentsprocessAn INS-to provide continuous measurements of图position, velocity, and attitudeRadio altimeterto provide continuousmeasurements of air vehicle height Above GroundLevel (AGL)An image capturing device (e. g. a video or Infra-Redcamera) and digitiser-to provide digitised images ofthe terrain beneath the air vehicleA computer running the algorithmsa database of linear features for an area covering theentire trajectory of the air vehicle for the currentIn order to achieve the afore-mentioned goals, howeverCVN was designed from the outset tomatch lines as vectors instead of pointsbe able to produce a measurement update to theFigure /-Creating CVN databa.ses from satellite imagerynavigation filter from only a single linear featurematchHowever, historical forms of line matching have suffereduse simple and fast image processing techniquesfrom high mission planning workload, due to theiroperate at least two, distinct, position fixingrequirement for(comparatively rare) intersecting linearalgorithms(designed to provide less of a compromisefeatures. They have also suffered from a lack ofbetween accuracy and robustness than singlerobustness, often brought about by the ambiguity of linearalgorithm techniquesfeaturesmaintain multiple, parallel ins bias hypotheses (inhe form of several Kalman Filters)These two failings, in conjunction with an out-datedconcern that databases would not be available. led to lineOf the above, probably the most significant in terms ofmatching systems losing favour, when all that wasenhanced robustness is the use of muliple ins biasrequired was a fresh approachhypotheses. The principle of their operation is as followsCONTINUOUS VISUAL NAVIGATIONWhen there is uncertainty about the best way to matchdetected features(from the captured image) and referenceThe Continuous Visual Navigation system [1 wasfeatures (from the database), Cvn maintains severalconceived by the navigation research group at Derahypotheses, cach in the form of an alternative Kalman(Farnborough) in conjunction with Hi-Q Systems LtdFilter. Subsequent fixes are then used to clarify which ofand has been funded by the United Kingdom Ministry ofthe hypotheses are unlikely to be true and these spurioushypotheses are rejected and ignored from then on. ThusCVN maintains a list of hypotheses, held in order of187likelihood. This list is constantly updated by morelive data from external sources and providing timelyinformation as the flight progresses. At all times, however,navigation data in a fast jet environment, is non-trivialthe system has an absolute ' hypothesis, which isHowever, over the last couple of years, this is exactlyoffered as the best estimate of position at that timewhat has been achieved with CvN. In addition furthertaken place in parallel with trials system2.1Primary Algorithmdevelopment in order to investigate potentiaimprovements to system performance. As part of thisCVn continuously propagates forward all active Kalmanwork, significant improvements to cvn performancefilters in order to react to changes in air vehiclewere demonstrated when a new approach to hypothesismovement. When any of the filters indicates that a linearscoring was implementedfeature will shortly be overflown, the following sequenceof events takes placeHypothesis scoring is the process by which the currenthypotheses are ranked in order of preference and iscvn predicts the time at which that lincar featuredistinct from covariance. At any point during operationshould be centrally located within the field of view ofCVNs best estimate of position is defined as thethe imaging sensor(currently a video camera).(Ahypothesis with the best score. Each hypothesis score islinear feature being centrally located within the Fovmodified according to information derived from eachof the sensor should correspond with the air vehicleiteration of the line matching processbeing directly above the feature.When this moment arrives it 'grabs' a digitised imageThe old hypothesis scoring approach was based on afrom the video camerameasure of filter consistency with some heuristics toThis digitised image is then filtered in order toreduce the score for erroneous or missed line matchesenhance linear featuresWhile effective in many circumstances, the consistencyThis enhanced image is then processed to identifymeasure could score inaccurate hypotheses highly as thelinear features and represent them in vector formlarge actual errors were still consistent withThe vector lines are transformed from image tohypothesis' estimate of that error. In areas of high featureground co-ordinates according to the height AGLambiguity, this would often lead to poor navigationmeasurement provided by the Radio Altimeter, andperformance. However, despite its crudity, it performedthe attitude information provided by the insremarkably well and saw Cvn through the feasibilityThe transformed line(s) are then matched withstudy phase.features in a local area of the database. in order toobtain an estimate of air vehicle position. If severalThe new hypothesis scoring approach is founded onplausible matches are found, then the same number oflikelihood. and is designed to avoid situations where acopies of each of the current Kalman Filters aredetected line matches extremely well with the wrongmade, and each match incorporated as a lineardatabase feature, as these occurrences corrupt cvnsmeasurement update into one of thesefilters. The approach computes the following likelihoodsa likelihood score is assigned to each KalmanFilter, and the few most likely Kalman Filters retainedThe likelihood that the line match is correctfor consideration on the next measurement updateThe likelihood that the line match is accidentalThe likelihood that image features have not been22Reversionary AlgorithmAs the name suggests, the reversionary algorithm wasIn areas of ambiguous linear features, there is a highinitially conceived as a recovery mode in case of primarylikelihood that good matches will be randomly achievedalgorithm failure. However, for improved performance, itfrom an incorrect hypothesis, and by comparing thenow runs in parallel with the Primary algorithmlikelihood that a line match is correct with the likelihoodthat it is random, incorrect hypotheses are less likely to beThe reversionary algorithm involves searching a largerpromotedarea of the database than is searched by the primaryIgorithm, and seeks to match a set of features that haveAs part of the preparation for cvn flight trials, atestalready been detected in captured imagesharness? was constructed. This harness uses recorded datafrom previous flight trials to simulate a real-time flightenvironment for CVN. This has proved an invaluable tool2.3Recent Improvements to CVNwhich has saved many hours of expensive flying time toThe process of taking a set of off-line experimentaldebug some tricky problems which only manifestalgorithms and turning them into a real-time system, usingthemselves in a dynamic environment. A significant188number of these problems concerned the handling ofexternal dataTRIALS INSTALLATIONThe cvn algorithms have been coded into software to runon a standard PC (currently 200MHz Pentium) runningcrosoft nt. In order to function in real-time. the Pcaccesses external data via the following interface cardsData Translation Dt3155 PC-compatible imagegrabber boardDDC 65539 PC-compatible MIL-STD-1553Binterface board (to access live INS and RadioAltimeter data)Datum-Bancomm PC03XT PC-compatible IRIG-Binterface board (to access GPS Utc current sourceof time).All hardware required for Cvn operation as well as theother navigation experiments, are carried in a converted1000 litre fuel tank(the ' pod), which is mounted beneaththe fuselage of the DERA Tornado(see Figure 2)Figure 3-Downward-looking video cameraThe Cvn PC Sits inside one segment of the pod (ssee那令gure 4)Figure 2- DERA Tornado and podhe cvn vidco camera views the ground throughcircular window on the underside of the pod(see Figure 3.)Figure 4-CVN PC inside podTORNADO 2000 FLIGHT TRIAL RESULTSCVN is envisaged to be used primarily as a reversionaryposition fixing system for when GPs data becomesunavailable. To ensure mission success. the missionplanner would be wise to adopt a pessimistic view ofwhen GPS will be jammed, and consequently ensure thatthe cvn database covers the whole of the mission profilefrom this point on. As stated earlier, this is not anexpensive or time consuming task189In order to model this scenario during multi-purpose flight200trials, on entry to the database area cvn is initialised with175a single kalman filter that has the correct ins bias values(according to the blended INS/GPS message from an onboard Honeywell H764G system). Subsequently, CVNuses raw ins data to propagate all its filtersa75Flight 1,50c25The first flight trial of cvn in 2000 was intended as afirst airborne test of the latest improvements to cv0102030405060708090100Although cvn is primarily aimed at navigationElapsed Time(s)applications where a low grade INS will be used (e.gmissiles), on this flight CVN was provided with data fromFigure 7-CVNradial postion error at 218m AGLthe aircraft grade INS (Honeywell H764G)that is part of(Flight 1)the current pod installation. Although the H764G is anINS/GPS SyStem. it makes available the raw iNs dataIn the figures(and all similar ones in this paper, the INSwhich continuously drifts from the Gyro-Compassdrift for the period of Cvn operation (red/dark line)isalignment that takes place before take off. The raw INSdisplayed on the same axes. The ins drift is calculated asposition error is shown in Figure 5-H764G INS radialraw radial ins position error minus the value that it had atthe start of cvn operation. This 'drift' makes a fairposition error.comparison of INS and Cvn position error, since(asstated earlier) Cvn was initialised with an accurate2000estimate of the ins bias values (provided by reference to1750the H764G blended data) at the start of each pass over the1500database-to simulate operational conditions12501000△7504.2Flight 2500The final flight trial of cvn in 2000 represented the most250severe test for CVn, as the environment closely0approximated that of an operational missile. Instead of0120024003600480060007200being provided with slow drifting aircraft grade INs data,Elapsed Time Since CVn Turned On(s)CVN was instead provided with the output from theFigure 5-11764G INS radial position error(Flight 1)DERA navigation research group 's " TIGNS. TIGNSstands for Tightly-coupled closed-loop IMU GPSFigure 6 and Figure 7 show the position error achieved byNavigation System, which actually quite effectivelyCVn (green/lighter line) during passes over the databasedescribes what it is The imu that is used within TIGNsat 84m and 218m respectivelycan be changed quite easily, but for the 2000 Tornadoflight trials the imu used was a Litton ln2o0. The200LN200 is Litton's name for a family of sensors, quoted in175[2] as providi0150L125Accelerometer bias repeatability of 0. 2 to lmilli-g100a75Gyro bias repeatability of l to 10%/hr (lo)G50Testing in the DEra navigation research groups inertialtesting laboratory has shown that the ln200 within0102030405060708090100TIGNS performs towards the higher quality end of theElapsed Time(s)above ranges. The GPs within TioNs is an adaptation ofFigure 6-CvN radial position errorat 84m AGLthe raytheon 2515 system. TIGNS is also used to test(Flight 1)Transfer Alignment algorithms [3]Although also an inertial/GPS system (like the H764G),GPS updates (to the tigns Kalman Filter that modelsLN200 instrument errors) can be manually enabled and190disabled by the Tornado navigator. As part of the flight200plan, which incorporated a variety of navigation75experiments, the navigator disabled GPS fixes just as theaircraft first approached the cvn database The aircraftLthen passed into and out of the database area three times100with Gps fixes constantly disabled. Consequently, the875position error of TiGNS was considerably larger during50the second and third passes through the database than itwas during the first. Although CVn was still initialised25with an accurate estimate of the tigns bias at the start of0102030405060708090100each pass it had to deal with the rapid error growth due toElapsed Time(s)the time since the last gps fixFigure 10-CVNradial position error at 155m AGLAs an indication of the magnitude and rapidity of tins(Flight 2)position error growth when GPS fixes are disabled, Figure8 shows the TIGNS radial position error for the whole of200Flight 21759150100175000e15000075可5012500081000007500001020304050607080901005000025000Elapsed Time(s)Figure 11-CVN radial posilion error at 8m AGL0120024003600480060007200(Flight 2)Elapsed Time Since CVN Turned On(s)Figure&-TIGMS radial position error (Flight 2)5SUMMARYAs in the figures for Flight 1, Figure 9, Figure 10, andCVn operated reliably and never became ' lostFigure 11 show the Cvn radial position error on the samethroughout both flights. Thanks to the recentimprovements, CVNs performance was better than cveraxes as the "INS(in this case TIGNS after GPS fixing hasbefore. in addition a useful observation that can be madebeen disabled)from the above figures is that CVN's performance did noto British Crown Copyright 2001vary noticeably at aircraft heights between 84m and 218mPublished with the permission of the Dcfcncc Evaluation andAGLRcscarch agency on behalf of thc Controllor of HMSOHowever, the most significant result of the flight trials is200that CVN's accuracy was only degraded by approximatelyE17550%o when using missile grade data instead of aircraftb150grade data. This is despite tiGNs position error growth出125being approximately two orders of magnitude greater thanthat of the H764G. this is very encouraging75particularly as the Cvn improvement work is not yetG50complete250102030405060708090100Elapsed Time(s)Figure 9 -CVN radial position error at 214m AGL(Flight 2)191THE FUTURE OF CVNConsequently, adding Cvn to an integrated navigationsystem comprising INS, GPS, and TCN is a cost-effectiveCvn has come a long way in its short history, but there isway of significantly increasing the chance of missionsome work still to do to quantify and refine its performancesuccessThis work includes:REFERENCESExtension of Cvns databases to permit evaluation ofperformance over different terrain, duringmanoeuvres. for longer periods[R. J. Handley, J. P. Abbott, and C.R. SurawyContinuous visual Navigation An Evolution of sceneRefinement of cvn's measurement process toMatchingProceedings of the Institute of navigationncrease overall system robustnessNational Technical Meeting january 1998)Study of how CVN's multiple hypotheses can best beused in a centralised integrated navigation system[2]LN-210 Inertial Navigation Unit Technical DescriptionLitton Systems Inc, Document No. 20484B July 1993)In addition performance evaluation is required3 Paul D. Groves and Jonathan C. Haddock, An AllUsing more sophisticated Kalman Filter models inpurpose Rapid Transfer Alignment Algorithm Set,place of the existing simple modelProceedings of the Institute of Navigation NationalAt greater heights and with a large initial error.Technical Meeting(January 2001)Using an Infra-Red cameraAs part of the study into how CVNs multiple hypothesescan be used optimally in a centralised integratednavigation system, work is currently in progress withMatra Bae Dynamics to investigate methods forintegrating Cvn within a production, conventionallyarmed stand-off missile INS/GPS/TCN navigation systemCVn could significantly extend theof potentialflight profiles of such a system and significantly reducemission planning time.CONCLUSIONThere can be no doubt that an INS/GPS system, possiblywith a crpa. will form the basis of all future airbornemilitary navigation systems. However, the risk of jammingis now too great to ignore, and CrPa technology isexpensive and still cannot guarantee mission success. Forthis reason, one or more accurate reversionary positionfixing systems are neededTCN is a good place to start but the accuracy is not greateven over undulating terrain. Over flat terrain it is uselessHowever CVN:-provides high position accuracy,requires a low mission planning workload,works best over flat terrain (where GPS jammers aremost effective)is low costrequires minimal processing power,could use existing image sensors192