Context Fusion for Driveability Analysis

Transcription

1 Context Fsion for Drieability Analysis Fredrik Lantz, Ssanne Edlnd, Erland Jngert FOI (Swedish Defence Research Agency) Box 1165, S-%81 11 Linköping, Sweden {flantz, Abstract - Drieability analysis is a qite complex problem that for its soltion depends on seeral factors. One of these factors concerns the type of ehicle for which a drie-way shold be determined. Besides this, the terrain strctre, the type of egetation bt also the grond type and its conditions play important roles. Drieability analysis will conseqently inclde analysis of primarily geographical information and the otcome of this analysis can be sed to spport decision making in command and control systems. Howeer, qite often the reqired geographical information is represented in a resoltion that is either too low and/or is represented with a high degree of ncertainty that cannot be neglected. In this work, an approach to drieability analysis is presented in which geographical information is regarded as context information that eentally is fsed to generate paths, that may be driable for certain types of ehicles. This information is fsed by means of a knowledge-based techniqe that determines the drieability from a set of qalitatie drieability impact factors. 1Introdction Drieability analysis (also called trafficability analysis) of terrain and geographical data offers an important techniqe for decision spport for all kinds of moements in the terrain. This type of analysis is needed for the jdgement of possible moements of targets (target tracking) and for the planning of ftre moements. An important data sorce for sch an analysis is a digital terrain model [1], [2], which in this case is a high resoltion digital terrain model generated from laser radar. In [2] is a method for finding essential terrain objects sch as ditches and ridges described. The geographical data are generally represented as map data that in order to become sefl need to be in a high resoltion. Parts of this work has earlier been described in [3]. The research on drieability has been sbject to fairly intense stdies e.g. Donlon and Forbs [4] hae deeloped a domain theory for trafficability to partition regions according to some criterion. They discss both complex factor oerlay and combined obstacle /05/$20.00 oerlay as a means 2005 to doieee this. A complex factor oerlay partitions regions according to the type of terrain and combine them into areas of homogenos characteristics. A combined obstacle oerlay characterizes the terrain according to its effect on the ehiclar moement. Bonasso [5] presents a trafficability theory sing first-order predicate calcls to enable reasoning abot trafficable paths withot the need for a detailed terrain model where only a limited set of qalitatie ales are sed, which e.g. allows a desert to hae the ale SOFT gien to a rigidity attribte, and FREE to a forest density attribte. Mch existing work on trafficability has focsed on real-time applications, mainly path planning for nmanned grond ehicles (UGVs), which moe in nknown areas [6], [7], [8], [9], [10] and [11]. In the work by Johnson et. al. [12] and Krse et. al. [13] hyperspectral or mltispectral data are sed to classify areas according to their srface composition, bt with limited sccess. Krse et. al. improe the seflness of their approach by sing data fsion to identify areas which e.g. hae both high clay content and high slopes. Another fsion oriented approach is gien by Glinton et al. [16]. Saponas et. al. [15] se an object-oriented approach and cost fnctions to calclate bonding regions and shortest paths. The cost fnctions represent the effect of the terrain on the trael speed. Finally, Slocm et al. presents an approach based on a trafficability engine [14]. The rest of this paper is organized as follows. In Section 2 context fsion is defined. This section incldes also a definition of the problem addressed in this paper. A discssion of existing drieability impact factors is discssed in Section 3. The digital terrain model sed is described in Section 4. Drieability analysis, basically inclding the sed measres of drieability and the final cost fnction with its impact factors is discssed in Section 5. Finally some reslts are presented in Section 6 followed by the conclsions and some frther research topics in Section 7. 2 Context Fsion The main objectie of this work is to deelop a method for drieability analysis that determines whether a certain path throgh a terrain area is

2 possible to trael gien a certain type of ehicle. The reslt of this analysis shold be isalized on a map where the drieability is highlighted. To be sefl in real-world applications information from a large nmber of sorces is reqired. The focs in this work, howeer, is concerned with the oerall method and basic strctre of drieability. Still, in addition to the digital high resoltion data model other geographical information is sed as well, e.g. forest data inclding the density of the forest. Finally, information abot the ehicle is sed, both with respect to its properties (e.g. if it has wheels or tracks, its weight etc.), and its capabilities (e.g. how steep slopes it can climb). This problem trns ot to be complex where the arios types of information mst be fsed and presented to enable generation of complete drieability maps. Here context information refers to the geographical backgrond information as well as the terrain eleation data model. The different sorces of context information needs to be fsed with the ehicle information to create the desired map. The reslt of this fsion process ths itself proide a new context, and is therefore referred to as context fsion. To carry ot the fsion, the impact factors of all the arios types of context information mst first be determined and then, finally, the drieability cost is determined throgh a cost fnction. The impact factor of a particlar context data type refers to the factors in the context that impact drieabilty. 3 Drieability impact factors Drieability is a complicated matter which does not lend itself to simple soltions. It is affected by many factors, inclding the ehicle type, soil, weather, slope, etc. Althogh otside the scope of this work, a particlar problem is to collect all the reqired information and represent it in a reasonably high resoltion. Drieability is path dependent. It may e.g. be possible to drie down a slope, bt not the other way arond. Most terrain featres cannot be regarded in isolation. Below are some examples: - A road barrier, a large stone, or a tree can only be seen as obstacles if it is impossible to drie arond them. - A wide ditch may be seen as an obstacle to a drier who mst cross it, bt not to a drier who can drie inside or arond it, or se a bridge. - A single tree may not be an obstacle, bt a dense forest can be a great impediment. - A slope may not be too steep, bt if the soil rigidity is too yielding it may be an obstacle. Vice ersa, a md field may be drieable on a plane area, bt not at a slope. - A ditch may not be drieable if it is filled with water, nor if trees or other obstacles are present in its proximity. - Vegetation and crops hae an impact on the drieability of soils. E.g., grass and grain often improe the drieability, while ineyards decrease it. There are also other featres, which are not intitiely possible to represent in a geometric model, bt which may still affect the drieability. For example srface roghness or egetation may decrease the speed, damage the ehicle, or bring it to a stop. Both stickiness and sinking soil properties affect the drieability, bt possibly in different ways. Sometimes the effect is not strictly physical. Mine fields and radar installations may be examples of nonterrain featres which decrease the wish to drie throgh sch areas, whereas a forest may be a good place to hide. Strictly speaking, these aspects shold instead be dealt with in threat analysis, althogh they still desere to be mentioned in this context. Many terrain featres change their properties oer time. Forests are growing denser and higher, riers are bending, etc. Dring a war bridges, roads, etc., may be destroyed. Weather properties may affect the terrain properties, thereby affecting the drieability. For example, soil rigidity is weather dependent. Dring a dry period a md field may trn into a passable field. A cold winter may trn a rier, that is otherwise impossible to pass, into a frozen road. Howeer, it is not sfficient to know what the weather is like right now; historical ales mst be taken into accont as well. The lake will not be frozen simply becase the temperatre falls below -10 C right now; it mst hae been cold enogh for a longer period. Hard winds or floods may also affect the drieability. Example of ehicle properties inclde width, length, height, oerride diameter, maximm gap to traerse, grond clearance, maximm step, maximm gradient, maximm tilt, specific grond pressre, and maximm straddle. Properties sch as maximm speed is not of any direct interest to the drieability reasoning performed in this work, althogh it can be sed as a basis for drieability calclations [15]. Sometimes the terrain featres and the ehicles interact in nexpected ways. E.g., een if a ehicle can oerride small trees, the reslting pilep of egetation may be a too great obstacle for it to pass. This effect is greater for wheeled ehicles than for tanks [17]. The properties are also not constant for the same featre. A ditch or a road, etc., may for instance grow wider, or fork, which is de to changes across the space. To start with, it may not be necessary to know the exact ale of a property. It may be enogh to se qalitatie ales (e.g. width is eqal to large), which can later be transformed into a real ale, as is sggested by Bonasso [5], or by an interal as will be sggested in this paper. As a conseqence of the complexity of the problem, the complexity mst be redced with respect to basic assmptions and simplifications, which will be discssed frther sbseqently. 4 Digital terrain model The data sed for creation of the 3D terrain model is registered by a scanning airborne laser-radar called Top- Eye 1. Uncertainties of the terrain eleation is in this case mainly dependant on the sampling density, which aries depending on flight eleation and speed. The data sed in

3 this work has an aerage density between 0.3 and 0.4 m. In order to detect significant terrain featres, the original sensor data is firstly pre-processed to eliminate non-terrain data, e.g. trees and bildings. The pre-processing step prodces a srface defined on a rectanglar grid with 0.5 m between grid points. As a first step, the srface is partitioned reglarly into tiles, which is a sb srface that coers a sqare with sides 2 m. The set of tiles is called T. These tiles are sorted into a nmber of categories of tiles, the spatial categories. A spatial category is a set of tiles with similar featres, e.g. with similar inclination or with edges at similar locations. The categories is defined by sing a set of representatie tiles REP T. Three different types of representatie tiles are necessary to achiee a sfficiently accrate srface model. The most important of which can be described by rotations and translations of the basic category forms, shown in figre 4. In total, 115 spatial categories are sed. For frther characterization see [2]. A distance metric dist: T REP R + for comparing the similarity of a tile to a representatie is defined. A category [r], r REP is defined, sing the metric dist, as the set of all tiles that are more similar to r than to any other k REP. Figre 1: Basic category forms. Eery category corresponds to a niqe string, called a symbolic tile. A tile is compared with eery representatie and are sbstitted by the corresponding string. The tiles can be stored in a database in accordance to a relatiely simple strctre [1], [18] from which data can be efficiently accessed. 4.1 Finding and connecting segments A filter corresponds to a particlar seqence of connected symbolic tiles describing the featre that will be sbject to the search. The approach taken here is to se the specific and characteristic cross-sections, which exist for all terrain formations. For instance, a cross-section of a hill can first be described with an pward directed slope followed by a slope directed downwards. At a conceptal leel the cross-section of a terrain formation consist of a start segment, a possible middle segment and an end segment, see figre 2. The filters can be of different sizes depending on the reqested terrain featre. Conseqently, the search strategy mst be sfficiently flexible to accommodate these demands. It trns ot that a two leel search strategy is the most appropriate.as a reslt of the first matching step a large nmber of segments will be fond. To find ot 1. A property of TopEye AB, Sweden Figre 2: The filter strctre to be applied to the crosssection of a ditch. which of these segments that are part of the featre, adjacent segments of similar cross-sections mst first be connected.segments fond in step 1 are connected only if the segments midpoints are sfficiently close and if the orientation of the cross-section segment is similar enogh. Some methods for doing so is described in [2]. The filters are specified in a plain text file that is parsed by a Jaa-program and applied to the symbolic tile data file. A nmber of filter types for different featres hae been deeloped and tested. Among them can filters for determination of ditches, ridges, hill tops and flat areas inclding roads be mentioned. At the same time as determining whether an object is present or not, an initial estimate of featre area and some impact factors can be calclated. In particlar, slope angles of slopes, conexities and concaities, as well as eleation difference and width of concaities, fig 3. Apart from sering as impact factors for the drieability analysis, these ales are sed for post-detection reftation of featres to exclde small and shallow featres of negligible inflence. Algorithms for acqiring better estimates of impact factors are nder deelopment, some possibilities are gien in [3]. a) b) c) d) l hyp l h Figre 3: Method to get an initial estimate of featre width. a) The featre. b) Horizontal width (l h ). c) Vertical width (l ). d) Featre width w = l h * l / l HYP. 4.2 Geographic data w The major sorce of data besides the 3D-data is the real-estate map, which contains data abot the geographic classification of the coered area in scale 1: These data are organized in oerlays, where each oerlay represents a certain geo-class, e.g. bildings and grond classifications are fond in different oerlays. Apart from proiding the drieability analysis with necessary data, the map proides a highly strctred representation of an area of interest (AOI), which is adeqate for ser presentation and interaction. Conseqently, when fsing the map data with other sorces it is desirable to keep most map strctres that are consistent with l

4 the more recently collected data. Althogh small, the data ncertainty de to imperfections in the collection process can not be disregarded when fsion with other high-resoltion sorces is at hand. More importantly, generalizations de to aggregation, simplification, smoothing, exaggeration, displacement etc., [19], are typically made with conseqences for the locational accracy that is difficlt to assess. Problems concerning the relatie positions of the terrain featres erss the map featres ineitably arises, prodcing apparent contradictions. When sering as a basis to drieabilty analysis, another problem concerns the ery rogh, sometimes none, estimate gien for most impact factors, e.g. the width of roads or the density of trees in a forest. In the absence of sch ales either defalts mst be sed or experienced sers mst be conslted for more specific estimates. Een thogh the map is a general and important sorce of geo-class data there may be other sorces of sch data. For example [18], proides a way to locate roads, bildings and indiidal trees by sing the eleation and intensity retrns of the laser radar. Sch data will improe the accracy of some locational estimates as well as proide the drieability analysis with some of the otherwise absent ales of impact factors, bt will not be considered frther in this paper. 5 Drieability Analysis Drieability is a measre of the possibility of a certain ehicle to follow a path p = (p 1,p 2 ) from the start p 1 to the end p 2. There are off corse infinitely many sch paths between p 1 and p 2 and there are infinitely many start and end points in a gien AOI. Using a finite, bt high resoltion of 0.5 m still makes the analysis prohibitiely expensie in comptational terms. To redce the comptational complexity, as well as to lay a fondation for ser interaction and spatial reasoning, another approach is sggested. Instead of iewing the AOI as an image, the AOI shall be iewed as a set of objects. These objects hae attribtes and belong to different classes. Conseqently, in order to be able to calclate the drieability for a gien AOI, a commitment to which objects, classes and attribtes the drieability analysis recognizes has to be made. The objects of concern in this application will be called terrain objects (TO). The complete enmeration of the entities that constitte the terrain objects in this application differs depending on the aailable data sorces and their implicit ontology. Any physical map entity that has a spatial extension, inclding roads, bildings, fences and similar line objects that direct inflence drieability mst be considered terrain objects. Eery 3D terrain featre is also a terrain object. In analogy with the map, the reslt of applying a certain filter to the laser-radar data defines a separate oerlay for each different filter. The terrain objects from different oerlays mst then be fsed to constrct a releant basis for a drieability analysis, i.e. a drieability segmentation. The reqirements of sch terrain object fsion will be discssed below. Fsioned terrain objects will be called compond terrain objects (CTO). A sitable drieability segmentation proides a complete segmentation of the AOI, where each CTO belong to three different types of classes. The types of classes are 3D classes, coer classes and obstacle classes. 3D-classes = {Concae, Conex, Slope, Flat, Concae&Slope, Conex&Slope, Undetermined} Coer-classes = {Water, Mash, GrondVegetation, OpenGrond, Road} Obstacle-classes = {Bilding, Forest, NoObstacle, DenseUrbanArea, Remains, LineObstacle} Eidently, the Coer-classes and the Obstacle-classes contain (essentially) a sbset of the classes aailable from the map. These classes are the necessary and sfficient classes to express the map data complexity in the crrent ersion of the drieability analysis. Each classification inflences which calclations that are necessary and that which makes sense at all. The classes also differ in their defalt properties, in the way they are inflenced by weather and their compactness. Frthermore, the obstacle classes can be both elements that are considered single entities (Bilding, LineObstacle) and elements, which are considered aggregations of smaller obstacles (Forest, DenseUrbanArea, Remains). Bilding, LineObstacle and DenseUrbanArea are compact objects, whereas Forest and Remains are not. The 3D-class Undetermined mst be inclded as the terrain featre filtering can not be expected, as the map, to proide a complete segmentation of the AOI. 5.1 Path selection Drieability, as defined aboe, is a concept that is meaningless withot a path to refer to. In this work, the focs is on the analysis of meaningflly, i.e. qalitatiely, different paths that are representatie of an AOI. A path is qalitatiely different than another path if it traerses a different CTO or if the traersal of a certain CTO is made in a qalitatiely different way. The qalitatiely different ways of traersing a CTO depends on its classifications. For instance, it is only meaningfl to differentiate between paths going from p 1 to p 2 and from p 2 to p 1 if the area is sloping sbstantially. Also, it is only meaningfl to consider leaping oer an object if it is a concaity, not if it is a conexity. In this paper, howeer, the same type of paths are considered for eery CTO. In fact, the paths are identified with traersal directions. The possible directions are D = {E,NE,N,NW,W,SW,S,SE}, representing traelling east, north-east etc. The directions D can also be considered defalt paths that shold be sed for the 3D class Undetermined, i.e. if the CTO is labelled Undetermined, the qalitatiely different paths traersing the CTO are members of D. 5.2 Reqirements of Terrain Object Fsion As mentioned, the data can be iewed as organized by a set of oerlays, where each oerlay contains a partial segmentation of the AOI. The fsion task at hand is to proide the drieability analysis with a complete, non-conflicting

5 segmentation of the AOI, where all areas are CTOs. An approach to constrcting a drieability segmentation wold be to simply intersect all oerlays [4]. The problems that arise de to ncertainties and generalizations in the data, changes in the terrain and the fact that when some classifications coexist at a location a conflict is at hand. Examples of conflicts are CTOs classified as roads with crossing conexities or as concaity and water. These types of conflicts mst be detected and resoled if the desired segmentations are to be achieed. As sch, the terrain object fsion problem incldes association, change detection and decision fsion problems commonly encontered in spatial data fsion [20]. An important consideration in this context is that the desired representation is not a drieability segmentation where the location of the objects from seeral sorces are combined to gie a more accrate locational estimate. Rather, the map representation shold be kept for the earlier mentioned reasons. In particlar, topological relations between TOs are of importance in the drieability segmentation. Een if not conflicting, withot proper consideration, the drieability segmentation may consist of nnecessary many, small CTOs which defeat the entire idea of redcing the AOI into a manageable nmber of homogeneos entities. In the crrent ersion, the map are assmed to be correct and accrate in all aspects. If a conflict with 3D-data occrs, the map has precedence. 5.3 Impact Factor Attribtes Apart from determining which class of TO that are present at a certain location, some attribtes, i.e. the impact factors, of the objects mst be estimated as well. In this case conexity gap width, maximm slope angle, minimm coer rigidity, obstacle rigidity, maximm obstacle rigidity and minimm obstacle distance. The intended physical interpretation is indicated in table 1. Rigidity is sed as a collection term for material properties, [5]. The corresponding attribtes for the ehicles are gap capability, slope p capability, slope down capability, grond pressre, force limit and ehicle width. The se of maximm slope angle, minimm coer rigidity, maximm obstacle rigidity and minimm obstacle distance is connected to the fact that the slopes, coers and aggregated objects hae different ales at different parts of the single path nder consideration. Howeer, the ale that determines the drieability of the path is the most disadantageos ale of the encontered ales at any part of the path. A path between p 1 and p 2 going throgh a solid rock obstacle is not driable not withstanding how easy the path part ntil encontering the rock may be. Table 1: Interpretations of arios impact factors. Impact Factor Interpretation conexity gap width maximm slope angle minimm coer rigidity obstacle rigidity Obios. The maximm angle of slope for all parts of a single path. The srface ability to withstand pressre from aboe. Measred in pressre nits. The obstacle ability to withstand force when being drien into. Measred in force nits. Table 1: Interpretations of arios impact factors. Impact Factor Interpretation maximm obstacle rigidity The ability to withstand force when being drien into. Releant for CTOs belonging to aggregated obstacle classes. minimm obstacle distance Releant for CTOs belonging to aggregated obstacle classes. 5.4 Impact Factor Variation As described earlier, drieability depends on a complex set of factors, many of which are difficlt to assess and for which data is freqently missing. The actal formlation of the conditions that mst be met for drieability is, howeer, not difficlt to express once the particlar case is determined. If it is the case that the path is going down a slope, throgh open grond and no obstacles are present, the slope angle mst be compared with the ehicle capacity in this regard and the rigidity of the grond mst be compared with the pressre to the grond exhorted by the ehicle. The reslt is tre or false and the drieability is ths determined. This conclsion is only appropriate if the CTO is completely niform oer the entire object in all properties, e.g. the object has the same maximm slope angle for all paths of the considered class. Clearly this is rarely the case. This is a case of spatial ariation for path s at indiidal TOs. Also, the performance characteristics of the ehicles of the same type is not niform, i.e. type ariation exists. Apart from the type ariation, the ehicle capacity is dependent on, for instance, the carried load and the maintenance of the ehicle, i.e. temporal ariation exists. Type and temporal ariation is typically a reslt of the qery of drieability not being specific enogh, i.e. the qery is pt in terms of the drieability of a certain type of ehicle at some nspecified time instant. In order to accommodate for the ariations exhibited by terrain objects, as well as by the ehicles, an extension to the approach mst be considered. Instead, consider the attribtes to be ariational qantities, i.e. the attribte ale is different at different parts of the CTO. For instance, the concaity gap width aries somewhat at different locations and the maximm slope angle aries at different parts of a slope. In this work, ariation will be described by an interal, [a,b], for eery impact factor. This is eqialent to saying that considering different parts of the CTO, the minimm ale of the impact factor is a and the maximm ale is b. The ariation of the maximm obstacle rigidity and the minimm obstacle distance connected with the aggregated classes are mainly connected to the fact that there are different types of smaller objects in the area and that the obstacle rigidity and obstacle distances aries between objects. The ariations described aboe can not, in general, be considered to describe the ncertainty of the tre ale of an impact factor. An estimate of the tre width of a ditch at a certain location can be sbject to ncertainty depending on the resoltion of the data etc., bt this is not what is described by an impact factor interal [a,b]. Neither shold that interal be interpreted as saying that the probability of the width at different locations is niformly distribted

6 oer the interal. As already mentioned, the drieability is a measre of the possibility of traelling a certain path. Hence, it is not concerned with the length of the path, or, gien a random choice of paths or ehicles, with the probability of sccess or with aerage performance. The choices are not random, bt made by skilled operators. Therefore, the statistics of random choices are ninteresting, only possibilities are needed. Still, a confidence measre that reflects the accracy of the data and of the estimation process is sefl to proide operators with an idea of the robstness of the analysis. Deciding that the ariation of a width of a ditch is [a,b] from measring two locations of a 100 m long ditch wold clearly jstify low confidence in sch a description. For the 3D data, the confidence is essentially connected to the density/resoltion of data after the grond filtering step, the properties of the grond filtering process and the age of the data. Crrently, no measre of confidence is calclated, bt all data necessary to do so is aailable. Expert sers or defalt reasoning can often proide a qalitatie ale for an impact factor or a ehicle capacity as an approximation. These qalitatie ales can be interpreted as interals as well, i.e. sing the ale SOFT for a rigidity attribte can be interpreted as saying that the rigidly is in some interal [a,b]. Using this interpretation, the same theory can be sed regardless of the sorce of the estimates. Hence, the absence of specific estimates as well as nspecifically made qeries and ariation in indiidals can be modelled sing interal aled ariables. 5.5 Impact Conditions As already mentioned, the conditions that need to be ealated to determine drieability are not complicated. De to the interpretation of impact factors and ehicle capacities as arying, the ealation of the conditions are, howeer, not eident. As mentioned, the ariation of ehicle capacities are both class ariation and temporal ariation, i.e. the entity for which drieability is determined is a generic instance of a certain class of ehicles and the determination is alid in some normal, possibly temporally changing, bt wide range of conditions. In accordance with the aboe terminology, the conditions to ealate are called impact conditions and the reslting ales are called impact costs. Animpact condition is a conjnction of relations between impact factors and ehicle capacities of the type: (impact factor i relation ehicle capacity i ). The only relations needed are <= and >=. The possible relations between two interals [a,b] and [c,d] are 13 as described in [21]. In this case, six ot of these can be gien distinct interpretations regarding drieability, see figre 4 for the relation <=, as an impact factor and as a ehicle capacity. The interpretations of these interal relations are, in decreasing order of drieability: 1. All ehicles can pass at eery location of the CTO. 2. Some ehicles can pass at eery location, bt some ehicles only at some locations of the CTO. 3. All ehicles can pass at some locations and none can pass at all locations of the CTO. 4. Some ehicles can pass at eery location, some ehicles only at some locations and some ehicles can not pass at any location of the CTO. 5. Some ehicles can pass at some locations and some at no locations of the CTO. 6. No ehicle can pass at any location of the CTO. Ealating a single impact condition ths gies an impact cost in {1, 2, 3, 4, 5, 6}. Combining the ales of impact costs from conjnctions of impact conditions are done by taking the maximm of the ales from the indiidal conditions. The crrently sed CTO classifications and the impact conditions that are ealated can be seen in table 2. The last entry in the table is a conjnction of the second and fifth entries. Figre 4: The interpretation of <= when considering and as interals. For any of the traersal directions D crrently nder consideration, the correspondence of that direction with a certain way to traerse the CTO mst be determined, e.g. whether the direction corresponds to going down or p a slope. This is soled by approximating the orientations of the TOs as one of the directions in D. The 3D-class Undetermined mst be inclded for the areas where no niform featre can be determined by the filtering process. A CTO belonging to this class has e.g. different slope angles associated for the 8 directions in D. 3D class 1 2 coer class obstacle class Concae Any NoObstacle Conex, Any NoObstacle Concae, Conex&Sl ope, Concae&Sl ope Slope Any NoObstacle way of traersing across on srface on srface impact condition (gap width <= gap capability) (maximm slope angle <= slope p capability) & (maximm slope angle <= slope down capability) & (minimm srface rigidity >= pressre) (maximm slope angle <= slope p/ down capability) & (minimm srface rigidity >= grond pressre)

7 3D class coer class grond pressre) Table 2: Some of the impact conditions for certain CTO classifications. 6 Reslts obstacle class Any Any Bilding, LineObstacle Flat Any Aggregated Not Flat Any Aggregated way of traersing throgh obstacles on srface and/ or throgh obstacles on srface and/ or throgh obstacles impact condition (obstacle rigidity <= ehicle force limit) (maximm obstacle rigidity <= ehicle force) & (minimm obstacle distance >= ehicle width) (maximm obstacle rigidity <= ehicle force) & (minimm obstacle distance >= ehicle width) & (maximm slope angle <= slope p capability) & (maximm slope angle <= slope down capability) & (minimm srface rigidity >= An experimental tool for drieabilty analysis has been deeloped. Two types of ehicles hae been sed in the experiments so far. One wheeled troop carrier and one tank. The capacities of these ehicles are sometimes ery difficlt to obtain. This fact can be modelled by sing a qalitatie approximation. The ale of a drieability analysis is of corse dependent on the estimation accracy of ehicle capacity bt, as shown in for instance [4] and [5], qalitatiely aled approximations can still be a road and some open areas in connection with these. The road is colored green in the figres becase of its low drieability cost. Following the road on each side is ditches. Other concaities and slopes can also be seen. Visalization of the drieability is handled by color coding of the tiles belonging to the terrain objects. In this case impact costs = 6 are colored red and impact costs = 1 are colored green. All other impact costs are white. Figre 6: The drieability of the carrier in direction north-east. Figre 7: The drieability of a tank. The direction of highest impact cost is selected for display. 7 Conclsions and Frther Work alable. Figre 5: The drieability of a carrier. The direction of highest impact cost is selected for display. As examples of drieability, a sqare AOI of sides 200 m is gien in figres 5, 6 and 7. The area contains Many data sorces are necessary if sccessfl drieability determination shall be possible. In particlar, 3D data in high resoltion is important to determine terrain featres and their attribtes. Fsion of context sorces and ehicle information is reqired to enable adeqate analysis of the properties of the context in which the ehicles moe, i.e. context fsion is needed. Reqirements on data sorces and processes are described. A formlation of drieability

8 analysis in terms of terrain objects, impact factors and impact conditions is presented. The sggested method acconts for the dependence on both path and ehicle characteristics. Most terrain featres and ehicles exhibits ariation that can not be neglected. An approach to handling this phenomenon is presented. Handling confidence measres in the estimate of sch ariational qalities is still to be resoled. A prime candidate for locational ncertainty representation in this context is rogh set theory, which is increasingly poplar in handling geographic information. References [1] Lantz, F., Jngert E., Dal aspects of mltiresoltion grid-based terrain data model with spplementary irreglar data points, Proceedings of the 3rd International Conference on Information Fsion (Fsion 2000), Paris, France, Jly 10-13, [2] Lantz, F., Jngert, E., Sjöall, M., Determination of Terrain Featres in a Terrain Model from Laser Radar Data, Proceedings of the ISPRS Working Grop III/3 Workshop on #-D Reconstrction from Airborne Laser scanner and InSAR Data, Dresden, Germany, October 8-10, 2003, pp [3] Edlnd, S., Drieability analysis sing a digital terrain model and map data, LITH-IDA-EX-04/031- SE, Linköpings Uniersitet, Linköping, Sweden, Mars, Also aailable as Technical Report, FOI- R SE, Department of Data and Information Fsion, Linköping, Sweden, May, [4] Donlon, J. J., Forbs, K. D., Using a geographic Information System for qalitatie spatial reasoning abot trafficability, Proceedings of the on Qalitatie Reasoning (QR 99), Volme 4364, [5] Bonasso, R. P., Towards a naie theory of trafficability, Proceeding of the Annal Conference on AI Systems in Goernment, [6] Broten, G. S., Digney, B. L., Perception for learned trafficability models, Proceedings of the SPIE, olme 4715, [7] Chatredi, P., Sng, E., Malcolm, A. A., Ibañez Gzmán, J., Real-time identification of driable areas in a semi-strctred terrain for an atonomos grond ehicle, Proceedings of the SPIE, olme 4364, [8] Digney, B. L., Learned trafficability models, Proceedings of the SPIE, olme 4364, [9] Dcksbry, P. G., Drieable region segmentation sing a Pearl Bayes network, In IEE Colloqim on Image Processing for Transport Applications, Digist No. 1993/236, [10] Grnes, A., Sherlock, J. F., Textre segmentation of defining driable regions, Proceedings of the British Machine Vision conference (BMV 90), [11] Jasiobedzki, P., Detecting driable floor regions, Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, [12] Johnson, A. J., Windesheim, E., Brockhas, J., Hyperspectral imagery for trafficability analysis, Proceedings of the IEEE Aerospace Conference, [13] Krse, F. A., Boardman, J. W., Lefkoff, A. B., Extraction of compositional information for trafficability mapping for hyperspectral data, In Algorithms for Mltispectral, Hyperspectral and Ultraspectral Imagery IV, [14] Slocm, K. R., Srd J. R., Sllian, J., Rdak, M., Colin N., Gates, C., Trafficability Analysis Engine, Cross Talk, The Jornal of Defense Software Engineering, Jne 2003, pp [15] Saponas, D., Kreitzberg, T., Johnson, M. L., Terrain trafficability model, In Military, Goernment and Aerospace Simlation, Proceedings of the 1996 Simlation Mlti Conference, [16] Glinton, R., Grindle, C., Giampapa, J., Lewis, M., Owens, S., Sycara, K., Terrain-based information fsion and inference, Proceedings of the 7th International Conference on Information Fsion (Fsion 04), Stockholm, Sweden, Jne 28-Jly 1, [17] Department of the Army, FM 5-33, US Army Field Manal, Terrain Analysis, URL seqrity.org/military/library/policy/ army/fm/5-33/defalt.htm, Jly [18] Elmqist, Jngert, Lantz, Persson, Söderman, Terrain Modelling and Analysis Using Laser Scanner Data, Proceedings of the Workshop on Land Srface Mapping and Characterization Using Laser Altimetry, Annapolis, Maryland, October 22-24, [19] Dnkars, M.,Mltiple representation databases for topographic information, Ph.D. Thesis TRITA- INFRA , Royal Institte of Technology, Stockholm, Sweden, December, [20] Waltz, E., The principles and Practice of Image and Spatial Data Fsion in Handbook of Mltisensor data fsion, Hall, D., Llinas, J., (Eds), CRC Press LLC, Boca Raton, Florida, USA, [21] Allen, J., F., Maintaining knowledge abot temporal interals, Commnications of the ACM 26, 1983, pp