A Mobile Mapping System Based on Phototriangulation

DEVELOPMENT OF A LOW COST MOBILE MAPPING SYSTEM:

A SOUTH AMERICAN EXPERIENCE

JOÃO FERNANDO C. DA SILVA and PAULO DE OLIVEIRA CAMARGO

São Paulo State University, Brazil

Abstract

This article presents the prototype of a low cost terrestrial MMS composed of a van, two digital video cameras, two GPS receivers, a notebook, and a sound frame synchronisation system. The imaging sensors are mounted as a stereo video camera on the top of the vehicle together with the GPS antennae. The

GPS receivers and the notebook are configured to record data referred to the vehicle position at a planned time interval. This position is subsequently transferred to the road images. This set of equipment and methods provides the opportunity to merge distinct techniques to make topographic maps and also to build georeferenced road image databases. Both vector maps and raster image databases when integrated appropriately can give spatial researchers and engineers a new technique whose application may realise better planning and analysis related to road environment. The experimental results proved that mobile mapping technology may be an effective approach to inspect road pavements, to map road marks and traffic signs, electric power poles, telephone booths, drain pipes, and many other applications important to the people's safety and welfare. There is already a small number of road images grabbed by the prototype as a consequence of its application in distinct projects. An efficient organisation of those images and the prompt access to them justify the need for building a georeferenced image database. By enlarging it, both at the hardware and software levels, it is possible to figure out for instance all the road environment available to scientists and engineers in their office computers.

INTRODUCTION There are situations when recently released technology is not easily accessible for reasons of high cost and due to political issues related to the acquisition of new technological equipment. In particular, INS (Inertial

Navigation System) technology has gained the attention of geospatial scientists and researchers over the past two decades. INS combined with GPS (Global Positioning System) is a very powerful technique for orienting imaging sensors in kinematics applications such as aerial photogrammetric or terrestrial mobile mapping projects. Although these technologies can provide both positioning and orientation for the vehicle and imaging sensors, in general, GPS is related to positioning and INS to the attitude.

Development of a mobile mapping system (MMS) prototype is a very difficult task in developing countries that are not designers of high technology equipment. In these countries, importation is the only way to access the technology. The difficulties to be overcome are mainly funding and the importing process when it comes to creating a complete configuration, comprising imaging sensors, GPS and INS integration.

Individually, the INS module is the most expensive. In the absence of INS, the phototriangulation technique based on the bundle block adjustment is a less expensive and reasonable method applied in the mainstream to overcome the sensor orientation problem, especially when it comes to deriving the camera’s attitude (Silva &

Oliveira, 1998).

Prior to the development of the mobile mapping prototype, a situation of collecting data on streets in order to make maps was simulated using an alternative methodology known as the photogrammetric traverse.

The idea behind that project was to provide the understanding of the map making process. Despite being an unconventional method for topographic mapping, the project highlighted the relationship between the stages that connect the main phases of data and information processing when it comes to the application of photogrammetric traverse. Details are given in Silva et al. (1999).

This article presents the prototype of a low cost terrestrial MMS composed of a van, two digital video cameras (Sony DSR200A), two GPS receivers - Ashtech Reliance (geodetic) and Garmin 12XL

(navigational) -, a notebook (Fujitsu Pentium 266MHz), and a sound frame synchronisation system (fig. 1).

The imaging sensors are mounted as a stereo video camera on the top of the vehicle together with the GPS antennae. The GPS receivers and the notebook are installed to record data referenced to the vehicle position at a previously planned time interval.

Besides the mobile unit, whose name in Portuguese is Unidade Móvel de Mapeamento Digital

(UMMD), in the Mobile Mapping Laboratory, there are computers used to download and process the digital images recorded in appropriate tapes (DVCAM format). The Mobile Mapping Laboratory represents the fixed segment where the digital images are downloaded to a hard disk. In this lab, vehicle positions, given by

GPS, are integrated to the digital images of a road using commercial and in-house software to complete the photogrammetric and cartographic phases. Both mobile (UMMD) and fixed (laboratory) segments comprise the first stage of a project titled Digital Mobile Mapping System recently concluded at the Department of

Cartography of São Paulo State University (unesp), campus of Presidente Prudente.

This set of equipment provides the opportunity to merge distinct techniques to make topographic maps and also to build georeferenced road image databases. Both vector maps and raster image databases when integrated appropriately can give spatial researchers and engineers a new technique whose application may realise better planning and analysis related to road environment. The experimental results proved that mobile mapping technology may be an effective approach to inspect road pavement, to map road marks and traffic signs, electric energy poles, telephone booths, drain pipes, and many other applications quite important to the people's safety and welfare.

The main qualities of such a terrestrial system are the speed, completeness, and effectiveness of acquiring field data in short time. The final output is easily integrated with Geographic Information Systems

(GIS). Cartography and GIS have increased their presence in modern economy as information technology becomes more important to business organisations. This new situation requires spatial information to create and to maintain databases. Modern data requirements necessitate distinct solutions for acquisition of spatial information. One solution is represented by mobile mapping systems (MMS). The growing use of GIS technology and the consequent necessity to update information increases the importance of an environment to handle spatial data and information that will be even more demanding. An efficient electronic deposit of a large quantity of road images obtained by the mobile system is envisioned, whereby an image database is required to manage the growing volume of data and images.

There are already a small amount of road images grabbed by the prototype MMS as a consequence of its application in research projects (Silva & Oliveira, 1998) and in graduation projects (Silva et al. 1999;

Silva et al. 2000). An efficient organisation of those images and prompt access to them justify the need for building a georeferenced image database. By enlarging it, both at the hardware and software levels, it is possible to figure out for instance all the road environment available to scientists and engineers in their office computers.

Developing countries need to improve their standards in quality of life. Maps can help especially when it comes to spatial information needs. Low cost systems are particularly interesting because they compromise requirements, results, and costs. Digital mapping solutions can satisfy both mapping requirements and reduce the technological gap experienced in most developing countries.

SYSTEM CALIBRATION

The MMS map making process is heavily dependent on photogrammetric operations like camera calibration, camera orientation, image co-ordinate measurement, and photogrammetric intersection. The importance of the imaging system is well known by the photogrammetric community. However, in developing countries access to adequate equipment due to costs or difficulties in getting technical assistance from foreign companies may determined the purchase decision. That was the case related to the projects reported in this article. Two digital video cameras (Sony DSR200A) were bought for their availability in

Brazilian market although they were not designed and built for photogrammetric or measuring purposes. As this type of camera is made for broadcast TV and video applications, one cannot expect a stable inner geometry. Although it is not meant for photogrammetry, it has useful capabilities such as sound, zoom, and an easy method of recording the images on DVCAM tapes.

Before going to the survey sites, a system calibration procedure has to be performed with both cameras mounted on the frame fixed on the top of the vehicle. The two digital video cameras are set together like a stereo camera in the normal case facing ahead with horizontal and parallel optical axes. Although set to manual and wide mode, when the cameras are turned on they automatically go to the shortest focal lengths

(nominal 5.9mm). Of course, the focal lengths vary around the nominal value every time the cameras are turned on. So it is necessary to assess the cameras’ interior orientation parameters. A calibration field was built with 54 targets on a wall. The targets had their co-ordinates computed by intersection after a base and the horizontal and vertical angles were determined using a total station (Sokkia SET5F). The vehicle, and hence the cameras, are positioned in front of the wall to take six images, these being two on the left, two straight in front, and two on the right. Images are taken with the left camera and three with the right in three different positions (fig. 2).

The images of the calibration test field are downloaded to a computer equipped with two hard disks: one IDE for the operating system and application software and the other (SCSI) for the non linear edition

(miroVIDEO DVTools for image capture and Adobe Première LE for image selection). One frame for each position is selected from a short strip (AVI format) and transformed to a bitmap image (BMP). FOTOCIC is the software developed with Borland Builder C/C++ 3.0 to measure image co-ordinates, which are then transformed to photocoordinates. The sensor size is 4.8 mm (h) x 3.6 mm (v) or 720 (h) x 480 (v) pixels, which means a pixel size equivalent to 6.7 m (h) x 7.5 m (v). Appropriate treatment is done to consider the aspect ratio. Table I shows four distinct inner parameter determinations, done in different epochs but under the same approximate geometrical conditions.

The variation is sufficiently significant to indicate that calibration is necessary for every new project.

The stereo base measurement was done using an invar tape that measured 0.94 m and this result was constrained in the self-calibration bundle block adjustment. Only four parameters (f, x0, y0, k1) were considered as the lens diameter is only 52 mm.

THE MMS MAP MAKING PROCESS

The traditional mapmaking process is based mainly on aerial photographs. Terrestrial photos proved to be effective only for very small sites or for projects with a few stereo-pairs. The integration of navigation and digital imaging sensors makes terrestrial photogrammetry effective for extensive road mapping projects.

Not only can vector maps be produced but also a myriad of data can be integrated in spatial databases such as road centrelines and signs, all visible features surroundings the roads, georeferenced raster images, and many others. Making maps with mobile mapping technologies is not intended as a tool to compete with the standard method. The best contribution of the MMS mapmaking process is its capability to be integrated or complemented with existing mapping technologies.

Field Surveying and Data Processing

The antenna of the rover receiver (Ashtech Reliance: carrier L1, code C/A) is fixed on the top of the vehicle half way between the two cameras. The antenna of the Ashtech Z-XII receiver (carriers L1, L2; codes

C/A, P, Y) is fixed on a reference station. The GPS signal acquisition rate is set to 1 second for both receivers. The synchronisation of the video cameras is started with the image acquisition by remote control.

The cameras' synchronisation to the GPS signal is done using a sound signal emitted by a Fujitsu laptop

(Pentium 266 MHz) when a special designed C language software (Hasegawa et al., 1999) records the position (geographic co-ordinates) signal emitted by the Garmin 12XL receiver every two seconds and the corresponding laptop clock time. The sound signal is sent to both cameras using a sound box from where two cables are introduced in both cameras' microphone jacks. This apparatus produces frames marked with a noise pick in the sound track corresponding to the laptop time that had been aligned to the Reliance clock (in order to have the same time base) that, in conclusion, means the geographic location given by the Garmin receiver.

Stereo base advances are the distances travelled by the vehicle along the roads during the time interval (2 sec) to get the next GPS location signal (given by Garmin 12XL) corresponding to the cameras' positions. They are schematically illustrated in fig. 3 (E1D1, E2D2, etc). The stereo base advances are therefore dependent on the vehicle speed. In highways, at 60 km/h, the time interval is equivalent to 33.3 m and in urban roads, at 20 km/h, 2 sec are equivalent to 11.1 m. In general, the base intervals can not be much larger than 30 m due to accuracy requirements (the longer the distance from the object point to the camera, the worst the accuracy).

The WGS84 (World Geodetic System 1984) co-ordinates for the antenna’s positions are transformed to a regional datum, for instance, SAD69 (South American Datum 1969), and then to UTM (Universal

Transverse Mercator) co-ordinates. The GPS data are post processed using Reliance software based on the kinematic GPS method. Under normal conditions, which means absence of tall structures like buildings or dense canopy trees for instance, the data processing realises standard deviations for UTM co-ordinates of about 2 cm for E and N co-ordinates, and 5 cm for the h component. The computed positions of the moving antenna station (P1 and P2 in fig. 3) are extended to perspective centres (CP1 and CP2, fig. 3). The height information (h) is derived from the GPS data and can be transformed to orthometric heights if a (local) geoidal transformation model is available.

Considering fig. 4, the two perspective centre co-ordinates are determined from two consecutive

GPS antenna positions by assuming that the stereo base advance (P1P2) is a straight line, the antenna centre

(P2) and the two perspective centres (CP1, CP2) are collinear with P2 in the middle, and both straight lines are orthogonal at P2. As previously indicated the stereo base length (CP1CP2) is 0.94m. The distance P1P2 and the azimuth AzP1P2 are derived from the computed GPS co-ordinates P1 and P2. The corresponding covariance matrices are also computed by applying the covariance propagation law. This process is repeated to compute the XC, YC, ZC co-ordinates for all perspective centres.

The total time spent in acquiring images and GPS signals in the field is very short compared to the total amount of time to execute the whole mapping project in the lab, especially when it comes to the photogrammetric and cartographic operations. Although time consuming, the MMS mapmaking process is mainly performed in a lab or office.

PHOTOGRAMMETRIC OPERATIONS

This mapping procedure requires only two photogrammetric operations with the image pairs in order to map the features of interest: sensor orientation and object point determination. The former is required to compute the attitude of the image pairs. The latter computes the object point co-ordinates based on the photogrammetric intersection.

Sensor Orientation

Suppose a sequence of stereo images is taken by a pair of cameras from the top of a vehicle that moves along a street or road. In a MMS, the images are usually positioned by GPS and oriented by an INS

(Inertial Navigation System). INS data fills in the blank positions when GPS data are not available. This is quite common in urban areas or when a vehicle travels through tunnels or under overpasses for instance.

When one does not have the INS to substitute the GPS in order to position the camera/image, orientation can be provided by a bundle block adjustment or other sensors.

As shown in the schematic representation (fig. 3), an object point or detail on a road can be seen in two or three stereo bases and then in four or six images. Simple, double or even multiple photogrammetric intersection can compute the spatial object co-ordinates. Theoretically, the closest stereo base delivers the highest accuracy and the far base the lowest accuracy for an object point when the photogrammetric intersection is computed separately for each base. For practical purposes, the pass points are selected from the two closest bases (stereo pairs). The photocoordinates are measured by clicking the mouse, which means visually, with an estimated error about 1 pixel. Six to eight pass points are measured in each stereo pair with four points common to two consecutive pairs. Figure 5 illustrates the measurement procedure with two pairs. When the pair at the bottom is finished it gives place to the pair above in which place the next pair of the sequence is loaded and so on. A bundle block adjustment is computed in order to estimate the attitude for both cameras at each station.

Although it is possible to compute the parameters of the whole sequence of stereo pairs, it is preferable to adjust all the interior, exterior, and object parameters involved only in a limited extension. This is due to the weak geometry as the pass points are not well distributed for the whole image and also the image-object vectors may have significantly different scales. There is no need for approximate object point co-ordinates because the bundle block adjustment is preceded by a preparation routine that computes all the object co-ordinates based on the photocoordinates and the approximate exterior orientation.

Photogrammetric Intersection

The road scene components like poles, trees, telephone booths, mail boxes, garbage cans, bus stops, traffic signs, the centreline, the construction and sidewalk lines, and the land lot (parcel) limits can have their spatial co-ordinates (UTM system) computed by photogrammetric intersection. The corresponding object point co-ordinates are computed from the twelve exterior orientation parameters and the four collinearity equations without any approximation. That is possible because some parameters are grouped together after some algebraic manipulation so that it results in a linear system of four equations with three unknowns (E, N, h).

Cartographic Design and Drafting

With MicroStation SE, all the intersected point and line features that are recorded in layers are edited and represented in a road map that includes contour lines interpolated from the orthometric heights.

The road and construction alignment points are constrained by a straight-line equation. In the road curves, an arc equation constrains the road alignment points. A topographic surface is interpolated to generate contour lines using Surfer software based on the mobile antenna heights.

The accuracy of the final product is estimated by comparison of the UTM map co-ordinates and

GPS ground co-ordinates (transformed to UTM) of the checkpoints. A trend and accuracy analysis, based on the methodology proposed by Galo & Camargo (1994), is performed with the computed E and N discrepancies and standard deviations. The trend analysis is based on the null hypothesis (Ho:  average = 0;

H1:  average  0) whose acceptation or rejection is controlled by the computed Student t statistics compared

to the theoretical tn;.

EXPERIMENTAL PROJECTS

Three experimental projects were conducted with the prototype system and derived data. The first project treated a sequence of street images and was concluded with production of a 1:2000 topographic map.

The second one investigated the mapping process of highways using a long sequence of road images and was finished with a 1:10000 topographic map. The last project focused on the development of a georeferenced road image database as an attempt to face the problem of organising the amount of images acquired in the previous projects.

Urban Road Mapping

The selected site for the mobile surveying and mapping project was composed of three urban blocks in the suburban area of the city of Presidente Prudente. The total time spent in acquiring images and GPS signals in the selected area was less than five minutes. However, it took around 20 minutes since the beginning of the section in order to solve the GPS ambiguities. WGS84 (World Geodetic System 1984) co- ordinates were transformed to SAD69 (South American Datum 1969) and then to UTM (Universal

Transverse Mercator,  = 51° W). The GPS data were post processed using Reliance software based on the kinematics GPS method and it released the following standard deviations for UTM co-ordinates of 57 stations: 1.7cm (E), 2.0cm (N), and 5.1cm (h). The computed positions of the 57 stations (P1 and P2 in fig. 3) were extended to the 114 perspective centres (CP1 and CP2, fig. 3) that correspond to the 57 stereo pairs. The height information was derived as described in section 3.1 using a local surface transformation model

(Oliveira, 1998). All the antenna stations and the point and line co-ordinates were positioned in a hybrid system formed by UTM (E,N) co-ordinates and orthometric heights (H). The hybrid system was considered orthogonal because of the small size of the field area and then the computed approximate object co-ordinates were introduced directly in the collinearity equations for the phototriangulation. The stereo base advances were planned to be around 20 m at maximum in the middle of the blocks and about 5 m when turning around the corners. This constrained the vehicle speed from about 36 km/h (10 m/s) to 9 km/h (2.5 m/s), respectively. In reality, the base advances in the middle of the blocks were about 15m.

The whole project was split into six distinct smaller blocks. Figure 5 shows that there were two blocks running north-south direction and four blocks east-west. For each of the six streets there were 19 images on average in each of the six split photo triangulation adjustments. With the perspective centre co- ordinates constrained to a weight of 10,000 (equivalent to a standard deviation equal to 0.01m), the computed average of the standard deviations was around 2' for  (pitch) and  (tilt or convergence) and 2° for  (roll along the road). The estimated standard deviations for the pass point object co-ordinates were about 1.5 m.

These results appear to be far from those given by classical aerial and terrestrial triangulation blocks but a direct comparison is not worthwhile because the network designs are pretty different. Despite this difference, it is important to suggest a few reasons to explain the results: the weak geometry of the photogrammetric traverse and the difficulties to choose pass points, both already mentioned by Silva & Oliveira (1998) and

Silva et al. (1999), and also the absence of automatic methods for point and line measurements.

With MicroStation SE, all the point and line features that were recorded in layers were edited and represented in a street map that includes contour lines interpolated from the orthometric heights. A straight- line equation constrained the road and construction alignments. At the street corners, an arc equation constrained the road alignment points. A topographic surface was interpolated to generate 1m contour lines using Surfer software based on the 57 mobile stations. Figure 6 shows the final street map at the scale 1:2000 entirely made from image and position data collected by the UMMD and processed in the Mobile Mapping

Laboratory. In fact, the cartographic original was made on an A3 size paper so that fig. 6 is a scale reduction.

The map accuracy was estimated by comparison of the UTM map co-ordinates and GPS ground co-ordinates

(transformed to UTM) of the 29 checkpoints. The average discrepancies () were -0.064m and -0.212m for

the E and N co-ordinates respectively. The corresponding root mean square errors (rmse or ) were 0.938m

and 0.811m. The trend analysis is based on the null hypothesis (Ho:  average = 0; H1:  average  0) whose acceptance or rejection is controlled by the computed Student t statistics – 0.36 m (E) and 1.41 m (N) – as compared to 1.7 m, the theoretical t28;5%. Table II shows the statistics.

According to the Brazilian cartographic law, the standard error (SE) in class A maps is equal to 0.3 mmM (M is the modulus or map scale factor), in class B SE is 0.5 mmM, and in class C SE is 0.6 mmM.

2 2 Considering these map accuracy classes, the null hypothesis ( comp   28;10%) is not rejected for class C maps at 10% statistical confidence level. In other words, 90% of all point co-ordinates extracted from the map are expected to have a planimetric error lesser than 2 m (1.0 mmM).

Highway Mapping

Approximately ten kilometers of highways nearby Presidente Prudente were surveyed with UMMD.

Both GPS data and video images were processed and transformed to a 1:10,000 topographic map. The field work and the lab processing were executed similarly to the urban road mapping project, except that the vehicle speed ranged from approximately 35 to 45 km/h (9.7 to 12.5 m/s). This means a stereo base advance ranging from 19.5 to 25.0 m. In the lab, the image co-ordinate measurement used a semiautomatic area correlation technique. The operator pointed the mouse in the left image and clicked in the surroundings of the homologous point in the right image. The corresponding point was selected by the highest correlation coefficient. This procedure reduced the time for image co-ordinate acquisition by one third.

As an INS was not available for UMMD, a bundle block phototriangulation was performed to compute the exterior orientation for the sequence of image pairs. The selection of pass points was more difficult than that in the urban road project because along the highways there were few well defined points.

Table III summarises the statistics from the highway mapping project. The last three columns show the average standard deviations for the attitude of the image pairs. 475 image pairs were measured and triangulated in 40 blocks. The average accuracy was about 38' and 20' for  and , and 2.36° for the angle .

The interest points like highway signs had their UTM co-ordinates computed by the linear photogrammetric intersection as described earlier.

Using MicroStation SE, all the extracted point and line features were edited and represented in a road map (fig. 7). When viewed on the computer screen, a click on a represented feature makes possible the visualisation of the image associated to that feature (fig. 8). Contour lines were interpolated from the orthometric heights of the moving GPS antenna. They were saved as a layer and they are shown only in a strip of about 30m that contains the road. A straight-line equation constrained the road alignment. In the curves, an arc equation constrained the road alignment points. The map accuracy was estimated by comparison of the UTM map co-ordinates and GPS ground co-ordinates (transformed to UTM) of the 14 checkpoints. The average discrepancies () were –0.313 m and –0.466 m for the E and N co-ordinates

respectively. The corresponding root mean square errors (rmse or ) were 1.729 m and 2.796 m. The trend analysis is based again on the null hypothesis (Ho:  average = 0; H1:  average  0) rejection criteria is given by the computed Student t statistics: -0.892 m (E) and -1.042 m (N) both smaller than 1.35, the theoretical t13;5%. Table II shows also the statistics for the highway project.

The accuracy of the 1:10,000 road map entirely made with the low cost MMS is compatible to the best quality map standard (class A) according to the Brazilian law.

Georeferenced Road Image Database

BIG stands for Banco de Imagens Georreferenciadas in Portuguese. It is the georeferenced road image database built with images obtained by UMMD during the surveyed urban road and highway projects.

The stored images may be completely oriented, case in which the full orientation (perspective centre co- ordinates and attitude) is known and stored in the image database. Also, either the position or the attitude might be known. The former case is preferable since it is possible to derive the attitude. Bundle block phototriangulation based on end pairs with their perspective centres weighted can solve the case when exterior orientation data are unknown, as described earlier. In any case, no ground control point is planned.

There is also the possibility of deriving the azimuth and the pitch from the perspective centre co-ordinates; roll can be neglected in some applications.

Builder C/C++ is the environment programming language used to develop BIG, which has modules to register in a cadastre, insert, modify, and delete records; queries and reports are also possible. The module of cadastre is responsible for the organisation of the information like project, client, city, neighbourhood, urban road or highway and the video camera. Each item has tools to verify and validate the data. All data registered in the cadastre module are stored in distinct tables; the written software manages them. After the above data have been registered, the images can be inserted. The image is the fundamental data of the database. There are restrictions to insert an image without having inserted the primary data (project, client, etc); it is not possible either to insert information related to a project without any image related to it.

Modification is another important module that practically can be considered an extension of the insertion module. In modification, there are the tools to update data. Depending on the application, this module can be very useful, especially in projects of temporal analysis. The exclusion module is intended to delete or erase the unnecessary or outdated information from the application or user’s point of view. Besides the exclusion of the data in the tables, it also excludes the images associated to the records. Data management is the most complex section of the software because it works with the manipulation of distinct tables, queries, and structures, besides the control of all data and images stored in the database. Practically, queries and/or reports present information from the image database. The queries provide users with visualisation of a certain group of information related to a specific point of interest. As already mentioned, an image in the database has to have all the information regarding the project, address, client, date and time of acquisition, environmental conditions; position, type and parameters of the camera, in addition to the exterior orientation parameters in a global reference system. Therefore any of the above items can be queried. The queries were designed to enclose the largest amount of data. They can be done by project, client, road (address), date, and by image. In order to have the requested information on paper in addition to the video screen, a report covers the same items as in the queries. Two report types were designed: simple and complete. In the former case, it prints general information with respect to the image while the complete type presents all items (localisation, project, client, exterior orientation, etc).

The efficiency of BIG relies only on the quality of the software. A great difficulty is the amount of hard disk space to store the image data. As the technology advances, this problem will be overcome in the near future. For the time being, this drawback has been bypassed by transforming the original 1.01 Mb bitmaps (BMP) to compressed JPEG 0.81 Mb images, which means an average saving of 12.5 times or 92% less disk space without any significant loss of image quality.

Among many applications, BIG can be used by municipal administrators to localise, visualise, and manage spatial information referred to a street or road, a land parcel, or a construction. BIG is interesting to the private and public organisations dealing with the planning, construction, maintenance, and use of highways and railways. Organisations responsible for the public utility services like water and sewer, electric power, telecommunications, garbage collection, traffic and transportation engineering, public health and safety, to say a few, can use BIG in their everyday operation. The image database can finally be used to update or to map urban details (like poles, manholes, trees, etc) by means of the photogrammetric intersection. A detailed description of BIG is given by Silva et al. (2001).

CONCLUSION AND RECOMMENDATION FOR FUTURE WORK The executed projects showed the potential of a low cost mobile mapping technology for mapping urban roads and highways. Bundle block phototriangulation is an option to circumvent the absence of inertial systems in order to orient the images. The workflow is heavily dependent on manual and visual operating work. This means a short time stay in the field and a much longer time in the lab. In the near future, new projects will rely on automated data acquisition both in the mobile unit and in the lab segment. The expected consequence is an improvement in data quality and faster processing and then a better accuracy of final product, which can be either a digital map, or an image database or both.

This paper also demonstrated the status of the development of a georeferenced image database named BIG to integrate a terrestrial digital mobile mapping system. It also mentioned some of the possible applications. The next phases of BIG are its integration with digital (raster and vector) maps and the improvement of query and report tools. It is envisioned that urban and road administrators can benefit from

BIG as its images can be integrated to dedicated GIS solutions.

Future work will concentrate on photovariant bundle block adjustment, enhancements to BIG, navigation by image analysis, and linear features to orient aerial and orbital images.

Photovariant Bundle Block Adjustment

A new project is under development to verify the effective contribution of the photo variant bundle block adjustment for the image orientation and for the determination of the object co-ordinates. This method will consider a set of interior orientation parameters for every image. Moniwa (1977) selected ten parameters per image, namely f, for focal length; x0, y0 for the principal point displacement; k1, k2, k3 for the symmetrical radial lens distortion; P1, P2 for the tangential lens distortion; and A, B for the affinity distortion. Experiments will indicate what group of parameters will be considered for every image in order to model the variant inner orientation considering the specific phototriangulation application of a pair of non photogrammetric video cameras.

Enhancements to BIG In the years to come, it is intended to integrate raster and vector data with other databases. The video clips of the streets and highways can be stored in a general-purpose image database. Extracted still images can be georeferenced by one of the methods mentioned above (direct-INS/GPS or indirect/phototriangulation) and then be stored in a georeferenced road image database which can incorporate vector maps or spatial (3D) geographic information systems or virtual reality applications (games or urban, architectural, engineering applications). This can be important for many engineering fields such as automobile navigation.

Navigation by Image Analysis

A road image sequence can be processed in order to help the navigation problem. Even with the development of the integration of INS and GPS methods for navigation purposes, image sequence processing can be an approach based solely on the images themselves independent of the external information. Of course, image navigation itself will not be able to estimate accurately the position and orientation of the camera (vehicle) as the digital image processing (DIP) errors will accumulate like inertial methods. The integration INS/GPS/DIP seems to be a very powerful and reliable method of navigation enriched by vector maps and image databases (Tao et al., 1999).

Linear Features

In theory, the absence of INS can be mitigated through the use of phototriangulation. However, point extraction from successive image pairs is not an easy task. The extraction of the abundant linear features along the roads may help the sensor orientation and navigation problem. Linear feature tracking in stereo image sequences for motion parameter estimation has proved to be effective in mobile mapping technology environment (Habib, 1994). The addition of linear constraints to the stereo video camera system self- calibration can be investigated. Orbital and aerial images can be georeferenced with the help of UMMD that can map the roads’ and rivers' trajectories. This will help to detect changes, to update, and to revise maps and databases. REFERENCES

GALO, M. & CAMARGO, P. O., 1994. Utilização do GPS no Controle de Qualidade de Cartas. Congresso

Brasileiro de Cadastro Técnico Multifinalitário, 1:41-8. (Portuguese).

HABIB, A. F., 1994. Estimation of Motion Parameters for Stereo-Image Sequences Using Data Association of

Linear Features. The Ohio State University: Ph. D. Dissertation. 129p.

HASEGAWA, J. K., GALO, M., MONICO, J. F .G. & IMAI, N. N., 1999. Sistema de Localização e Navegação

Apoiado por GPS. Congresso Brasileiro de Cartografia, 19(2):6p (CD-ROM). (Portuguese).

MONIWA, H., 1977. Analytical Photogrammetric System with Self-Calibration and its Applications. The

University of New Brunswick: Ph. D. Dissertation. 106p.

OLIVEIRA, K. C. L., 1998. Modelo Geoidal para a Região de Presidente Prudente. Faculdade de Ciências e

Tecnologia, Universidade Estadual Paulista. 46p. (Portuguese).

SILVA, A. R., BATISTA, J. C., OLIVEIRA, R. A. CAMARGO, P. O. & SILVA, J. F. C., 1999. Surveying and Mapping of Urban Streets by Photogrammetric Traverse. International Workshop on Mobile Mapping Technology,

Bangkok. Proceedings, 32(2W10): p. 5.A.5.1-4.

SILVA, J. F. C., CAMARGO, P. O., OLIVEIRA, R. A., GALLIS, R. B. A., GUARDIA, M. C., REISS, M. L. L. & SILVA,

R. A. C., 2000. A Street Map Built by a Mobile Mapping System. International Congress of ISPRS, 19.

Amsterdam. Proceedings, CD-ROM 1, Book 2 (Commission II), p. 506-13.

SILVA, J. F. C. & OLIVEIRA, R. A., 1998. Triangulation of a Sequence of Terrestrial Digital Images.

International Symposium - Data Integration: Systems and Techniques, Cambridge, UK. Proceedings, ISPRS,

32(2):273-7. SILVA, J. F. C., OLIVEIRA, R. & GALLIS, R., 2001. Georeferenced Road Image Database. International

Symposium on Mobile Mapping Technology. Cairo, Egypt. Proceedings, ISPRS. 3(11):8p. (CD-ROM).

TAO, C., CHAPMAN, M. A., EL-SHEIMY, N. & CHAPLIN, B., 1999. Towards Automated Processing of Mobile

Mapping Image Sequences. International Workshop on Mobile Mapping Technology. Bangkok, Thailand.

Proceedings, ISPRS, 32(2W10):5.A.5.1-4.

Acknoledgement: The authors thank to Fapesp (Fundação de Amparo à Pesquisa do Estado de São Paulo) whose grants (97/11049-6, 99/00807-2, 99/04831-5) made this work possible. They also thank to Dr. Mike

Stewart of Curtin University, Perth, Australia, for his comments. Fig. 1. Mobile Mapping Unit with the digital video cameras and the GPS antenna Fig. 2. Convergent calibration method. E 1 E 2 E 3 E 4

D 1 D 2 D 3 D 4 .Q

Fig. 3. Mapping the object points in a sequence of image pairs. Fig. 4. Two consecutive GPS antenna positions and the two perspective centres. Fig. 5. Interface to measure point and linear features for bundle block and intersection. Fig. 6. 1:2000 street map (reduced) made with data acquired by UMMD. EA

Fig. 7. Reduced scale (original 1:10,000) road map made with data acquired by UMMD. Electric power pole in the map linked to the image

Electric power pole in the image.

Fig. 8. A road map feature linked to the correspondent image. Table I. Four inner camera parameter determination projects.

Sony DSR200A I – First Calibration II – Second III - Urban Road IV - Highway

Project Calibration Project Project Project L: left; R: right P SD P SD P SD P SD L f (mm) 5.891 0.002 6.058 0.078 6.873 0.098 6.035 0.043

L x0 (mm) -0.132 0.002 0.008 0.031 0.553 0.002 -0.069 0.031

L y0 (mm) -0.001 0.002 0.004 0.030 0.077 0.036 -0.028 0.030 L k1 (mm-2) -0.00297 0.00016 -0.00398 0.00185 0.05000 0.003 -0.00872 0.0027 R f (mm) 6.041 0.002 6.000 0.056 6.406 0.069 5.981 0.030

R x0 (mm) 0.035 0.002 0.172 0.082 0.215 0.099 -0.052 0.084

R y0 (mm) 0.055 0.002 0.155 0.055 0.010 0.099 -0.080 0.056 R k1 (mm-2) -0.00615 0.00024 -0.00708 0.00231 0.00500 0.003 -0.01700 0.0065 L: left camera; R: right camera; P: parameter; SD: standard deviation Table II. Trend analysis and accuracy statistics of the urban road and the highway projects.

2 2 2 2 Parameter  (m)  (m) t sample t28;5%  A  B  C  28;10% Urban Road E -0.064 0.938 0.36 1.7 136.95 49.3 34.23 37.92 N -0.212 0.811 1.41 1.7 102.21 36.79 25.55 37.92 2 2 2 2 Parameter  (m)  (m) t sample t13;5%  A  B  C  13;10% Highway E -0.313 1.729 -0.892 1.35 4.99 1.80 1.25 19.81 N -0.466 2.796 -1.042 1.35 8.08 2.91 2.02 19.81 Table III. Average standard deviation of the estimated attitude by bundle block phototriangulation

Road Ext (m) blocks pairs pass II / IV IV / III V / IV SD  SD  SD  points (m) (I) (II) (III) (IV) (V) (VI) (VII) (VIII) (IX) (X) (XI) MG 1,140 6 56 313 20.4 9.3 5.9 1°16'42" 22'36" 3°37'24" RT 3,770 14 188 1042 20.1 13.4 5.5 20'32" 18'33" 2°04'17" JB 2,145 9 110 594 19.5 12.2 5.4 19'34" 20'03" 1°14'43" EA 3,075 11 121 649 25.4 11.0 5.4 35'48" 19'19" 2°30'21" Total 10,130 40 475 2,598 21.3 11.9 5.5 0°38'09" 0°20'08" 2°21'41" Roads: MG - Manoel Goulart; RT - Raposo Tavares; JB - Julio Budiski; EA - Amity.