DIC ATEF A. ELASSAL*

Raytheon CompanyAlexandria, Va. 2230-1-

Data Edit Using ModularComputer ProgramsThe success of very large simultaneous photogrammetricsolutions depends on an effective means for editing the input data.

INTRODUCTION

I T IS DIFFICULT TO CONCEIVE a successfulAnalytical Photogrammetric Data Pro­

cessing System without an effective dataediting package being associated with it. Thisfact becomes more evident when viewed inthe ligh t of the presen t days tendency towardsthe simultaneous processing of large photo­gram metric nets.

In order to arrive at a convenient form forsuch a data editing package, the general prob­lem of data editing is reviewed, a logical de­velopment of some possible solutions is at­tempted, and the specific structure of a pro­posed data editing system is outlined. Finally,an illustrative example of a possible solutionto one of the proposed system components isdemonstrated.

REVIEW OF THE DATA EDITDIGPROBLEM

In the early development of most of themodern treatmen ts of the problem of analyti­cal photogrammetry, data editing procedureswere approached as a natural extension of theleast-squares treatment. The computed resid­uals in a problem were usually edited bycomparing them to a certain multiple of thea posteriori statistical estimate of their stan­dard deviations. The theoretical validity ofsuch treatment, and the apparent ease of im­plemen ti ng it, produced a general concensusof its approval. The approach, as it stands,required the arrival to a least squares solutionbefore any data editing could be performed.

It was not until the later development of

• Presented at the Semi-Annual Convention ofthe American Society of Photogrammetry, St.Louis, Mo., October 1967, under the title, "The useof modular computer programs to edit data for verylarge simultaneous photogrammetric solutions."

large analytical data processing systems thatthe problems resul ti ng from mul ti-source datacollection ystems and the associated inter­face problems became eviden t. Errors causedby these problems are usually gross in nature.Failure to identify them could result in in­stability of the least squares iterative pro­cedure preventing its successful conclusion. Itthen follows that screening of these gross er­rors must necessarily precede the simultan­eous treatmen t of the data.

Manual edit procedures used as part ofquality control schemes could be injected atstrategic points of the data collection andpreparation system. This will undoubtedlyreduce the frequency of the occurrence ofblunderous errors but it will not eliminatethem. It is then paramount that an auto­matic edit procedure should be attempted.This could only be achieved, with the re-

DR. ATEF A. ELASSAL

1180

D.\TA EDIT USING i\\ODCLAR COMPUTER I'ROGRAl',,\S 1181

quired degree of infallibility through the useof com pu ter con trolled operations.

The question is then, "How to bring thepower of the computer to tackle efficientlythe problem of data edi ting?".

PROBLEM DEFINITION

I n order to define our objectives clearly, wewill first classify accidental errors in the basicobservations into two classes:

a. Errors which exceed a certain multiple of the apriori estirnate of the standard deviation of an ob­served quantity. These errors are ot such magni­tude that they cannot be explained as a result ofthe stochastic sampling process. ]-1 uman blundersand machine malfunctions are the largest con­tributors to this type of error. It is also this kindof error that causes the greatest difficulty inpractice. The reason could be understood by

error requires a rigorous statistical treatment. Atest of significance which could be built aroundthe least-squares adjustment procedure is usuallyused. The implementation of such tests does notrepresent any particular data-processing prob­lems and, thus, will not be treated in this paper.

We will now proceed to examine methods todeal wi th the fi rst type of error, described aboveunder paragraph a. The data processing aspectsof the problem will be treated in detail with noreference to any specific mathematical formula­tions that might be used to perform the edit.

CO~[PUTER ApPROACH TO THE PROBLEM

A photogrammetric situation could be re­constructed through the application of anordered set of mathematical operations. Eachone of these operations constrains the data insome way in order to satisfy a set of physicallyknown geometrical relations. Examples of

ABSTRACT: The practical usefulness of an analytical photogrammetric dataprocessing system is highly dependent on an effective data editing package beingassociated with it. The programming complexity of such a system is greatly en­hanced through the use of modular computer programs. The technique makes useof (L logical modulation of the photogrammetric reconstruction process to its basicccmponents. Errors detected in the data during the execution of the d~fferent re­construction steps are properly indexed. A logical interpretation of these indicescan then be used to identify data errors and their possible causes. The proposedsystem structure makes it particularly useful for on-line data collection andverification operations.

examining the nature of the mathematical modelswhich describe most of the photogrammetricsituations. These models are usually non-linearin the unknown parameters. As least-squaresadjustment could be only performed on linearsystems, linear approximations of those modelsare used instead. The linear approximations areusually obtained through a Taylor expansion ofthe original mathematical model around a pointin the n-space defIned by the problem's param­eters. The expansion is accomplished by ne­glecting second and higher order terms in theparameter residuals. This can be valid only if thepoint around which the expansion was carried isclose enough to the model's solution. The failureof this condition due to the existance of blundersin the data could result in the instability of thesubsequent least-squares adjustment of thederived linear model. The result of this is thefailure of the solution to converge. It is quitedifficult to state in general the size of the errorsthat will result in such instability. The geometri­cal situation and the degree of redundancy in theobservations playa decisive role in determiningthe error size that can be tolerated. Fortuntaely,the majority of photogralllmetric systems re­Ilects a strong geometry by design. A fair degreeof redundancy in these situations makes it pos­sible to tolera te a ra ther large size of errors in thedata without upsetting the stability of the leastsquares solution.

b. Errors which are within the probable accuracylimits of the data cot!ection system. This type of

such operations in the classical triangulationcase are:

• Reduction of the plate measurements to aform which reflects the principle of centralprojection.

• A reconstruction of the internal geometry ofthe photogrammetric net through the processof relative orientation.

• Scale transfer between different parts of thereconstruction.

• Formation of a photogrammetric model ofthe photographed object by simultaneousintersection of conjugate rays.

• Absolute orientation of the photogrammetricmodel in order to express the results in thedesired reference system.

Failure of any of these operations or the de­tection of excessive residuals will point tothe existence of errors in this part of the datathat is being operated on. The nature of theoperation being performed will also providethe clue to the possible causes of these errors.Error detection and isolation procedurescould be built in these operations. Each oper­ation must properly index that part of thedata found in error so that it will be identi­fiable to any subsequent operation. Becausecertai n types of errors are u niden ti fiable

1182 PHOTOGRAMMETRIC ENGINEERING

(masked) to some operations, the sequenceof application is important and could be usedto the advantage of increasing the effective­ness of the data edi t process.

The data edit system must provide for theapplications of any set of operations in anydesired sequence. This degree of flexibility isimposed on the system for a number of rea­sons:

* Prohlems to be handled will reflect a widevariety of geometrical configurations.* In the case of hybrid systems, provisionsmust be made to add or delete photographicor control information.* In case of on-line data-collection and verifica­tion, the system must be capable of handlingunscheduled accumulation of any type ofdata.

This required degree of flexibility could beachieved by employing modular program­ming technique.

1\ JODULAR PROGRMD11 KG

1t is possible to break down any data pro­cessing system into a number of logically in­dependen t opera tions. Each one of these oper­ations transforms the problem data to ahigher level of processing until the completeprocessing is achieved. A program is thus apreconceived assembly of a number of suchoperations. To facilitate the programmingand assembly of these operations, they areusually programmed in a form of modules.From the programming point of view, themodules must have identical form. Thismeans that the sequence of operations whicheffect a transfer of control to and from amodule must be identical.

Responsibility of controlling the assemblyand execution of these modules resides with aprogram which is usually termed the Super­visor. The Supervisor could assemble andexecute a program phase consisting of one ormore modules which may be arranged in anydesirable sequence.

The logic involved in controlling the relayof the required data to an arbitrary arrange­ment of the modules could be extremely com­plex. A great deal of this complexity can beresoh'ed by standardizing the data formatand by only allowing a program module tochange the data contents leaving its standardformat intact. However, each module shouldbe responsible to check the data in order tofind out whether or not it meets the module'sminimum requirements. The concept of datamanipulation and formating is termed DataManagemen t.

DATA MANAGEMENT

A large percentage of programming timeand effort are spent in setting up the mechan­ics of data storage and retrieval. This is par­ticularly true in photogrammetric applica­tions where both the amount and diversity ofthe data handled are usually extensive. Agreat deal of the efficiency of any data pro­cessing system depends on how data is manip­ulated between the different program mod­ules. One convenient way of keeping track ofdata storage and retrie\'al is through the con­cept of data sets. A data set is defined as anamed collection of data. Plate measurementsare an example of such a data set. The dataset in this case could be composed of imageiden tifications, plate coordi nates, covariancematrices of plate coordinates and a set of in­dices that indicates the data status. Theplate identification could serve as the dataset name. A number of data sets could be filedin a Catalog. Each entry in the Catalog con­tains a data set name and an indication of itslocation. The location indicator contains boththe storage device and where the data setresides on this device. The collection of datasets and their catalogs are defined as a dataVolume. 1n photogram metric applications,separate data volu mes could be used for platemeasu remen ts, camera station parameters,and ground control. It is to be mentioned thatthe cross reference between the di fferen tvolumes of data is usually carried throughcamera-station and point identifications. Itshould be also noticed that the status of eachpiece of data is pro\'ided by an indexing sys­tem which is part of the corresponding datavolume.

The index of an element of data must beable to indicate the set of conditions that theelement could assume. For instance, the indexof an image-point coordinates may indicatewhether the point is active or has beendropped. The same index can also reveal theprogram module which rejected that point.I n the case of the camera sta tion parameters,the index system could also reflect the type ofcoordinate system in which they are expressedA translation of the index system at any timewill provide a complete picture of the datastatus. Storage could be employed most effi­ciently if special number systems were usedfor each type of indices. The basis for thesenumbering systems will be equal to the vari­ous conditions the corresponding piece of datamay assume. Thus, a binary number systemwill be used to reflect a binary condi tion, andso on.

DAT.\ EDIT CST TG MODl'LAR COMPI'TER PROGR:HIS 1183

DArn TERMINlIoLS

r-----'I

I IL .J

r--------,

II I

L -.JSYSTEM SUPEIlVISOR

FIG. 1. Proposed organization chart for data varification and fding operation system.

The concept of data management as givenabo\'e is intended to give a general idea of theproblem extent. The details are far morecomplicated and allow the system designerample room for a great deal of creativity. Thepoint to be emphasized here is that, once adata managemen t system is established, theprogram modules using it must not be allowedto change its form. This arrangement willmake possible an arbitrary sequence of appli­ca tion of the progra m mod ules.

The system could be designed to operatein one of two modes:

Off-Line Operations. This mode of operationmay be used if data verification is to be per­formed after the process of data collection iscompleted. This will be necessary if a computeris not accessible to handle the data during its ac­cUllltl1ation.

The module-operating sequence can he gen­erated by the supervisor program or by the useras part of his input. Program-generated sequencewill be possible only in case of regular photo­graphic coverage "'here logical procedures couldbe employed to generate the se'luence.

The results of this operating procedure will bea complete diagnosis of the verified data pointingthe errors and their possible causes.

On-Line Operations. This type of operatingprocedure will fIt the needs of fairly large organi­zation where the corresponding large volume ofda ta ha nd led req uires efficien t use of its com­puting facilities. The system will be operating ona time sharing basis with different terminalsfeeding the computer, possibly in a simultaneousfashion. Data-fde updating and error detectionwill then be conveniently kept in step with thedata acquisition process. Error rectificationcould be achieved as the data is accumulated.

Figure 1 shows the proposed system's organiza­tional chart.

lLLUSTRATlVE EXA\IPLE

The principles stated here were attemptedon an experimental scale in a program forcantilever extension developed at Auto­metric/ Raytheon. The program re\'olvesaround two progra m mod ules; rela ti ve orien­tation and scale restraint. A wide \'ariety ofoperating options were built in the programin order to gi\'e the user a greater control overthe data Row. These options in terms of dataediting are:

AUTOMATIC MODE

1n this mode, the program performs dataediting automatically, guided with controlparameters gi ven to it by the user. Theseparameters, augmented with others which areinternally generated, provide the norms forthe data rejection process. For instance, inthe relative orientation process, the user pro­vides the program with a rejection criteriafor plate coordinate residuals. The user alsosupplies the program with the minimum de­grees of freedom that have to be retained inthe system. The plate coordinate rejectioncriterion EI will reRect the user's estimate ofthe maximum acceptable residuals based onhis evaluation of the material, equipmentand procedure used in collecting the data. An­other plate coordinate rejection criteria E2 isa certain multiple of the a posteriori estimateof the plate coordinate standard deviation.

llR4 PHOTOG RA,I ~I ETRIC I·;:-':GI:-': EERT:\G

TABLE l. [MAGE COOIWI1'IATES OF P0I1'ITS SHo\\'1'I IN FIGt;I~E 2

Point Left Nigh!

If)X (mm.) )I (m·m.) X (mnl.) )I ("/1111'1.. )

UA 0.000 110.000* -105.000 100.000UB 50.000 100.000 -55.000 100.000llC 100.000 100.000 -5.000 100.000:\rA 0.000 0.000 -105.000 0.000:\IB 50.000 0.000 -55.000 0.0001vrc 100.000 0.000 -5.000 0.000LA 0.000 -100.000 -lOS .000 -100.000LB 50.000 -100.000 -55.000 -100.000LC 100.000 -100.000 -5.000 -100.000

* Correct Value= 100.000Principal Distance =150.000 ITII11.

This value is computed by the program foreach attempted relative orientation solution.Plate residuals are first checked against E,. l[

any rejections are detected under this test, itwill be limited to those points with the highestresiduals provided that the requirement forminimum degrees of freedom is satisfied. Thetest is always performed on all points in orderto provide means for re-including points thatwere previously rejected but later were foundto be acceptable. The data edit process con­tinues until a stable condition is reachedwhere no points are either rejected or in­cluded. \\'hen this happens, the program pro­ceeds to perform the same operations againusing E2 instead of E,. The program will shiftto manual mode if the conditiono of the editwere impossible to fulfill 01- where the numberof editing trials exceeds a preassigned num­ber.

MANUAL MODE

I n this mode the program reli nquishcs con­trol to the user at certai n decision poi n ts. Theuser can then relay instructions about what

course of action that should be taken throughthe combined uo·e of console typewriter andsense switches. The user can also change re­jection criterion and return to the automaticmode if he wishes.

The manual mode was found to be neces­sary in order to provide a means to deal withmarginal cases which require intelligent de­cisiono which cannot be built in the programlogic without undue complications. Thesecaseo arise primarily where weak geometry orlow degree of overdetermination exists.

\\'e will now pl-oceed to illustrate themechanics of the data verification as exer­cised by the program. A simple example of asi ngle stereomodel is used. (See Table 1, Fig­ure 2). The model simulates two verticalphotographs with a total of nine points intheir common o\-erlap. An error of 10.000mm. is introduced in the y-coordinate of onepoint. The program was supplied with thefollowing parameters:

Maximul11 allowable plate residuals 0.050 111m.!VI inimum degrees of freedom 2.Maximum allowable number of trials 5.

Three trials were required to isolate theerror. Table 2 and Figures 3a through 3c il­lustrate a computer printout of the datastatus after each trial. In Figure 3a, the plate

FIG. 3. Points rejected by computation afterfirst, second and third trials as shown in Table 2.(Figures 3a, 3b, and 3c are numbered left to right.)

A B C A B C A B C

M U X X X U ~ ~ x U ~ x x2 7

M x x x M x x x M x x x4 9 3

x x x x x x x x x5 8 6

lEFT

FIG. 2. Single stereoscopic model and arrangementof points associated with Tablet.

DATA EDIT USING :I'IODUL\R CO:\[PL;TER PROGRAMS 1185

TABLE 2. PRINTOUT OF PI.ATE RESroUALS FOR TlIREL:: TRIALS ASSOCIATED WITH FIC;URES .la, 3b, AND 3c

Plate Residuals of Model Left -Righi Ser No

ID DXl DYI DX2 DY2Type

VA 0.070 1.105 0.025 -1.192 1

UB -0.101 -1.598 -0.031 1.680 2

VC 0.025 0.394 0.008 -0.404 3

MA -0.037 -0.014 0.014 0.783 4

MB 0.000 0.007 -0.000 -0.007 5

:'vlC 0.041 0.827 -0.013 -0.820 6

LA 0.016 0.436 -0.024 -0.439 7

LB -0.001 -0.016 0.001 0.016 8

LC -0.016 -0.421 0.022 0.40-1 9

Weighted Sum of Squares=0.11617785E 02 Degrees of Freedom =4 Unit Standard Devia tion = 1. 704

VA 0.000 5.000 -0.000 -5.000 OUT 1

VB 0.000 0.000 -0.000 -0.000 OUT 2

UC 0.000 0.000 -0.000 -0.000 3

MA 0.000 0.000 -0.000 -0.000 4

i'l'fB 0.000 0.000 -0.000 -0.000 5

:VIC 0.000 0.000 -0.000 -0.000 6

LA 0.000 0.000 -0.000 -0.000 7

LB 0.000 0.000 -0.000 -0.000 8

LC 0.000 0.000 -0.000 -0.000 9

Weighted Sum of Squares=0.20761789E-16 Degrees of Freedol11=2 Unit Standard Oeviation = 0.000

VA 0.000 5.000 -0.000 -5.000 OUT I

B 0.000 0.000 -0.000 -0.000 2

VC 0.000 0.000 -0.000 -0.000 3

~IA 0.000 0.000 -0.000 -0.000 4

MB 0.000 0.000 -0000 -0.000 5

MC 0.000 0.000 -0.000 -0.000 6

LA 0.000 0.000 -0.000 -0.000 7

LB 0.000 0.000 -0.000 -0.000 8

LC 0.000 0.000 -0.000 -0.000 9

\\"eighted Slim of Squares=O.15813595-15 Degrees of Freedol11=3 Unit Standard Deviation=O.OOO

residuals did not meet the rejection criteria of.050 mm. The program then rejected the twopoints UA and UB which exhibit the largestresiduals and at the same time retaining therequired 2 degrees of freedom. The programafter the second trial (See Figure 3b) acceptsthe previously rejected point UB. Figure 3cprei:ents the last trial in which the residualsare found to meet all the data editing cri­terion.

CONCLUSIONS

The modulation of the photogrammetricreconstruction process could be effectivelyemployed to produce a highly flexible dataediting system. The structure of such a sys­tem makes it particularly useful for on-linedata collection and verification operations.The extent of such system could be alteredwi th relative ease to make it fit a wide rangeof organizational needs.