COLOR CONSTANCY FOR LANDMARK DETECTION IN OUTDOOR ENVIRONMENTS

Transcription

1 COLOR CONSTANCY FOR LANDMARK DETECTION IN OUTDOOR ENVIRONMENTS Eduardo Todt Carme Torras Institut de Robòtia i Inormàtia Industrial, CSIC-UPC, Barelona, Spain. todt@ieee.org, torras@iri.up.es Abstrat This work presents an evaluation o three olor onstany tehniques applied to a landmark detetion system designed or a walking robot, whih has to operate in unknown and unstrutured outdoor environments. The irst tehnique is the well-known image onversion to a hromatiity spae, and the seond tehnique is based on suessive lighting intensity and illuminant olor normalizations. Based on a dierential model o olor onstany, we propose the third tehnique, based on olor ratios, whih uniies the proesses o olor onstany and landmark detetion. The approah used to detet potential landmarks, whih is ommon to all evaluated systems, is based on visual salieny onepts using multisale olor opponent eatures to identiy salient regions in the images. These regions are seleted as landmark andidates, and they are urther haraterized by their eatures or identiiation and reognition. 1 Introdution In reent years the idea o using legged robots in environments where wheeled robots are not suitable is assuming inreasing importane due to their superiority in limbing obstales and keeping stability in irregular terrains, despite their greater omplexity. Motivated by these harateristis, whih are appropriate to our appliations in outdoor unstrutured environments, we are urrently developing a prototype o a legged mobile robot [15]. A undamental issue in this enterprise is the position estimation problem. Sine, with the use o internal measures, the estimation o position aumulates errors as the robot moves, espeially or walking mahines in uneven terrains, it is neessary to develop loation tehniques based on some sort o external sensory eedbak, independent o the traveled distane [2]. Similarly as in human and animal navigation, we reord visual reerenes and use loal and temporal relationships between these reerenes in order to identiy plaes in the world, and to plan and exeute paths between loations [8]. These reerenes are alled landmarks, whih are distintive entities that the robot an reognize whenever they are in its detetion range. In general the haraterization o landmarks in unstrutured outdoor environments through some set o geometri eatures is not appropriate, in opposition to the ase o indoor environments, sine requently the elements that ould be taken as landmarks do not have well-deined shape and ontours. With this assumption, landmark haraterizations based on olor and texture eatures are more reliable, beause they are independent rom shape. One o the key ators that makes the detetion and reognition o visual landmarks in outdoor environments a hallenging task is that aquired visual inormation is strongly dependent on lighting geometry (diretion and intensity o light soure) and illuminant olor (spetral power distribution), whih hange with sun position and atmospheri onditions. In order to overome these adversities, the original Red, Green, and Blue (RGB) olor omponents o aquired images are oten transormed to other olor spaes, in an attempt to redue the dependene on lighting geometry and illuminant olor [4][14]. This desired invariane o olor representation to general hanges in illumination is alled olor onstany. In a previous paper [13] we presented a system or natural landmark detetion in unstrutured outdoor environments based on the model o human visual salieny proposed by Itti, Koh and Niebur [7]. This system had already inorporated a simple olor

2 onstany mehanism, onverting the input RGB images to hromatiity images. Although this mehanism provided some invariane to the detetion o landmarks under light intensity hanges, our experimental results indiated that it was neessary to searh or a olor onstany method more stable to the broad illumination hanges observed in outdoor environments. Taking into aount that mobile robot navigation and loation tasks require real-time exeution, the searh or other olor onstany methods to improve the landmark detetion was done onsidering low omputational ost as an important requirement. It is also worthwhile to remark that here the goal is to ahieve a stable landmark detetion system, and not to reognize olors with high auray. A reent method or olor onstany, whih takes into aount both lighting intensity and illuminant olor was proposed by Finlayson, Shiele and Crowley [4]. By ombining this olor onstany method with olor histogram omparisons, they ahieved good eetiveness in objet reognition tasks. Beause o its independene o illuminant models, olor patterns, and similarity with the hromatiity spae rom the irst approah, their method was seleted or our system. Thus, we itted it to our landmark detetion system as an image preproessor and exeuted a new sequene o experiments. These experiments indiated improvements in the stability o landmark detetion, but at the expense o signiiant additional omputational ost, due to the iterative nature o the involved omputation. Also this method presented the drawbak o sensitivity to hanges in viewpoint and to the inlusion o new objets in the senes, beause o its dependene on the global olor omposition o the images. With the objetive o reduing these dependenies and keeping omputational ost low relative to the salieny detetion task, we designed a novel algorithm or visual salieny ombined with olor onstany. The proposed algorithm is inspired by the olor onstany model proposed by Gevers and Smeulders [5], based on the olor gradients between neighboring pixels. Our approah presents the advantage that the olor onstany is embedded in the visual salieny detetor, instead o ating as an image preproessor. This paper presents an evaluation o these three dierent approahes to olor onstany applied to our landmark detetion system. First o all, the landmark detetion system is desribed. Then, eah olor onstany method is explained, together with its onnetion to the landmark detetion system, and the respetive results are presented. Finally, all methods are disussed and ompared between themselves. Considering that the goal o this work is to ompare dierent olor onstany approahes applied to landmark detetion, visual salieny is omputed here only based on the olor inormation available, disregarding other relevant visual ues, like intensities and orientations, whih play an important role in the omplete visual salieny system [13]. 2 Landmark detetion based on visual salieny When there is no exat knowledge o what things in the environment an be used as landmarks or the loation o the mobile robot, some riterion is needed to deide whih regions in the images will be onsidered as potentially good landmarks. Current theories o human vision indiate that potential landmark loations ould be deteted by using a visual salieny mehanism [3]. It has also been observed that when human pereivers are trying to build or reover the desription o a sene, their attention is oused on speii relevant regions in the sene [6], whih reinores the idea o ombining visual salieny tehniques with landmark based loation tasks or mobile robots. In the late 19 th entury, the German physiologist Ewald Hering laid the oundations o the olor opponeny theory, whih sustains the existene o three opponent olor proesses in the visual human system, onstituted o red-green, yellow-blue and intensity (white-blak) hannels [14]. The visual salienies in an image are proportional to the enter-surround dierenes o the hannels responses, omputed by omparing visual ield regions with their respetive neighborhoods. For example, red stimuli surrounded by green stimuli are onsidered salient, and vie versa. The same onept is also valid or blue-yellow and intensity hannels. The greater the ontrast between the enter and surround regions, the greater the orresponding salieny. We adopt the Enroth-Cugell and Robson s model [3] o enter-surround opponenies, whih onsiders the eet o the light weighted aording to the distane to the enter o the reeptive ield by a dierene o Gaussian untions. To ompute the desired olor opponenies, the trihromati input olor spae is transormed to an opponent olor spae, where it is possible to easily determine the red-green and blue-yellow opponenies. Ater this transormation the image is omposed o our hannels, eah one orresponding to an opponent olor. The image is then represented at eight spatial resolutions, through the use o our multisale Gaussian pyramids, one or eah hannel. In these pyramids, eah level is obtained by a low-pass iltering operation on the preeding level, ollowed by a sub sampling o ator two in eah dimension. With this struture, the entersurround dierenes are omputed by single dierenes between orresponding pixels at ine and oarse sales within the pyramids [7]. A enter region orresponds to

3 a pixel at a ine sale, whereas its surround orresponds to the respetive pixel at a oarser sale. An important property o the pyramid implementation is that it ailitates opponent omputations at dierent resolutions. By omparing enters and surrounds at high-resolution levels, e.g., between levels 2 and 5, visual salienies o relative small targets are ound, while at lower resolution levels, e.g., levels 4 and 7, relatively large salient regions are ound. With this tehnique, it is possible to detet visual salient objets within a wide size range, or example, rom small stones to big trees. The visual salieny system builds a salieny map, whih represents the distintness o eah point in an image. In this map, those points orresponding to salient areas have large values, whereas non-salient areas are indiated by small values. The omputation o enter-surround dierenes at dierent sale ombinations (pyramid levels 2-5, 3-6, and 4-7) results in partial visual salieny maps. The resultant partial maps are ombined into a global visual salieny map, whih has the ues to identiy potential landmarks. This salieny map is represented at the resolution o level 1 o the pyramids, to redue omputational ost, sine this level has a quarter o the number o pixels o level 0, whih orresponds to the original image size. The image regions orresponding to the potential landmarks are subsequently analyzed to obtain visual signatures, apable o identiying them as an existing or a new landmark, but this is beyond the sope o this work. 3 Visual salieny with lighting intensity normalization The irst olor onstany method we onsider in this evaluation is the transormation o the RGB spae to hromatiity oordinates (rgb) [14]: R r = (1) R+ G+ B G g = (2) R+ G+ B B b = (3) R+ G+ B The olors represented in rgb oordinates are muh more stable to lighting hanges than those in the RGB spae, beause the light intensity omponent is removed rom eah pixel. However, they ail to be invariant under spetral power distribution hanges o the light soure, beause this type o perturbation aets the response o the RGB sensors in dierent proportions. Following Itti, Koh, and Niebur [7] we use the opponent olor spae with our olor omponents, red, green, blue and yellow (R, G, B, Y ), alulated rom the rgb hromatiity spae as ollows, taking only positive values: R = r ( g+ b)/2 (4) G = g ( r+ b)/2 (5) B' = b ( r+ g)/2 (6) Y = ( r+ g)/2 r g b (7) The resulting opponent olor image represented in R, G, B, and Y oordinates is then proessed with the visual salieny system desribed in the preeding setion. The partial salieny maps obtained rom eah enter-surround sale (2-5, 3-6, and 4-7) and opponent olor ombination (R -G, G -R, Y -B, and B -Y ) are normalized aording to their maxima and then added to orm the salieny map. There are several other deinitions or opponent olor spaes that ould be used instead o the ormulation o Itti, Koh, and Niebur [7]. For example, Swain and Ballard [11] propose similar deinitions, although without a term onsidering the absolute dierene between the R and G omponents to ompute the yellow, while others are based on logarithmi dierenes, like that in Finlayson and Barens [1]. Finally, some deinitions aim to obtain better deorrelation between the olor omponents, suh as those in Otha, Kanade and Sakai [10], Murrieta-Cid, Briot and Vandapel [9], and Tan and Kittler [12]. We tested all these deinitions, and ound the adopted ormulation better than the others or our system. It is worthwhile to observe that opponent olor spaes tend to deorrelate the RGB omponents, whih is a desirable harateristi or pattern reognition. We have ompiled the experimental results or three senarios subjet to dierent real illumination onditions in Figures 1-3. The results orresponding to the lighting intensity normalization just explained are shown in the seond olumns o suh igures. In Figure 1, it an be observed that the red roo (enter let in the image), the orange lowers (enter right), the brown bush (bottom let), and the yellow house (enter) are indiated as salient, but the dominant green areas appear also as salient, whih is not desired. Also the gravel path is salient, espeially in the irst image. The stability o the results to illumination hanges in these images is poor. In Figure 2, the red house (enter), the yellowish bushes (bottom let and right), and the trees at the horizon line are indiated as salient. These bushes and trees are very salient and tend to redue the relative salieny o the red house. Near the enter o the seond image it an be observed a spot o salieny orresponding to a red sign. The green areas in these images produe a salient bakground, like in Figure 1, beause at a large sale (pyramid levels 3-7) they are salient relative to the sky

4 and near ground. In these images the stability in salieny is also poor, partly beause o the inlusion o new elements in the observed senes due to the hanges in point o view, whih aet the relative salieny between all the elements in eah image. This eet is most notieable in the third image, where the yellow bush is the most salient element in the sene. In Figure 3, the green areas are dominant in the sene but they don t have opponeny to other regions o similar size, like in previous igures, avorably resulting in lower bakground salieny. The yellow lowers (bottom right) are deteted as salient, and the reddish tree is indiated as salient only in the irst and third images. In the seond image the RGB values o the reddish tree are too low to produe reliable hromatiity inormation, and thus the orresponding region is dismissed. In these images the stability o salieny is good, mainly due to a uniorm bakground. These results orroborate the need o looking or a olor onstany method that, when inorporated in our salieny detetion system, produes salieny maps more stable to illumination hanges. The problem o inluding new salient objets, or example due to point o view hanges, like in Figure 2, ould be partially avoided onsidering separately the salieny maps or the R-G and Y-B olor pairs. 4 Visual salieny with lighting intensity and illuminant olor normalization In order to overome the unavorable sensitivity to illumination hanges shown by the previous normalization, Finlayson, Shiele, and Crowley [4] proposed an algorithm or olor onstany alled omprehensive olor normalization, based on iterating two types o suessive olor normalizations. These normalizations are aimed at removing dependene on both lighting intensity and illuminant olor, in an alternate manner. The irst normalization type is the same as that deined by equations (1) through (3), transorming the image to hromatiity oordinates. The seond normalization type transorms eah pixel aording to the global mean value o the olor bands: 1 r r ' = * (8) 3 r 1 g g ' = * (9) 3 g 1 b b' = * (10) 3 b where r, g and b are the mean value o the red, green and blue bands in the whole image. The olor onstany proedure iteratively perorms these two types o normalization until the dissimilarity between two suessive resultant images is below an aeptane level. It is possible to demonstrate that the method onverges and provides unique results, assuming that the amera sensors have a very narrow bandwidth [4]. The resulting normalized image is then onverted to opponent olor spae, aording to equations (4) through (6), and, inally, it is applied as input to the same visual salieny system desribed beore. Experimental results with the visual salieny detetion system using omprehensive olor normalization are shown in the third olumns o Figures 1, 2, and 3. In Figure 1, it an be observed that the red roo (enter let in the image), the orange lowers (enter right), the brown bush (bottom let), and the yellow house (enter) are still indiated as salient, but the dominant green areas appear also as salient, whih is not desired. The gravel path presents more stability and lower salieny than in the ormer experiments with intensity normalization. The salieny obtained with omprehensive olor normalization is outstandingly more stable to the onsidered illumination hanges than with intensity normalization (seond olumn), although there is an exessive salieny o green regions near the horizon. In Figure 2, the red house (enter), the yellowish bushes (bottom let and right), and the trees at the horizon line are indiated as salient. Here also the bushes and trees have high salieny, reduing the relative salieny o the red house, partiularly in the third image. The spot o salieny orresponding to a red sign is also present near the enter o the seond image. In these images the salieny stability is enhaned ompared to the intensity normalization (seond olumn), although the same eet as beore o masking salient regions due to inlusion in the image o more salient elements is notied. Sine the omprehensive olor normalization uses averages o olor omponents over the entire image, the inlusion o new salient regions aets the stability o salieny even worse than in the ase o intensity normalization. This eet an be observed in the third image where the red house is no longer signiiantly salient. In Figure 3 the salieny obtained through omprehensive olor normalization is stable, but without signiiant improvement over the lighting intensity normalization.

5 a d b a d b a d b Figure 1: Visual salieny omputed or sene A or three dierent illumination onditions (dierent days and times within the day). Eah soure image (let olumn) was proessed with lighting intensity normalization (seond olumn), lighting intensity and illuminant olor normalization (third olumn), and olor ratios (ourth olumn). The whiter regions indiate the more salient parts deteted. In the soure images the ollowing things are indiated: (a) red roo, (b) orange lowers, () brown bush, and (d) yellow house. g e g e h g e g e Figure 2: Visual salieny omputed or sene B or our dierent illumination onditions. Note that there are slight hanges in perspetive between the images. Eah soure image (let olumn) was proessed with lighting intensity normalization (seond olumn), lighting intensity and illuminant olor normalization (third olumn), and olor ratios (ourth olumn). The whiter regions indiate the more salient parts deteted. In the soure images the ollowing things are indiated: (e) red house, () yellowish bushes, (g) trees, and (h) red sign. The other signs, similar in shape as the red one, are white.

6 j i j i j i Figure 3: Visual salieny omputed or sene C or three dierent illumination onditions. Eah soure image (let olumn) was proessed with lighting intensity normalization (seond olumn), lighting intensity and illuminant olor normalization (third olumn), and olor ratios (ourth olumn). The whiter regions indiate the more salient parts deteted. In the soure images the ollowing things are indiated: (i) yellow lowers, and (j) reddish tree. 5 A new approah: visual salieny using olor ratios Gevers and Smeulders [5] show that several olor representation spaes, like hromatiity rgb, saturation S and hue H, the CIE standard set o primary olors XYZ, and pereptual uniorm spaes U*V*W* and L*a*b*, an be invariant to hanges in illumination intensity, but not to hanges in illuminant olor. These authors propose a new olor model, m 1 m 2 m 3, independent o the illuminant olor and based on the olor ratio between neighboring image loations. Assuming that the soure RGB omes rom narrow band sensors and that the olor o illumination is loally onstant, they demonstrate that their model is invariant to hanges in illumination intensity and olor, and also to hanges o viewpoint, objet geometry and illumination diretion. They deine the values o m 1 m 2 m 3 or eah pixel loation as: R G m1 = (11) G x1 R x2 R B m2 = (12) B R G B m3 = (13) B G where x 1 and x 2 are neighboring pixels. The authors implement the omputation o neighboring pixels ratios through the determination o gradients in the logarithmi spae ln(r/g), ln(r/b) and ln(g/b). The maxima o the gradients are deteted using a Canny edge detetor, resulting in a ontour image, that presents good olor onstany properties. From the deinition o m 1, m 2 and m 3, we derived the idea o generalizing the onept o gradient between neighboring pixels to that o enter-surround opposition. Under this approah, the x 1 pixels are taken to mean enter regions and the x 2 pixels orrespond to surround regions. Then, we deine the enter-surround olor opponenies based on olor ratios: RG RG s = (14) RG s RG s 1 GR = = (15) RG RG s BY s BY = BY s (16) BY s 1 YB = = BY BY (17) s where R, G, B, and Y are red, green, blue and yellow enter or surround regions, aording to the subsript. The RG opponeny orresponds to a visual ield that is exited by red stimuli in the enter and by green stimuli in the surround, and inhibited by red stimuli in the surround or green stimuli in the enter. The GR orresponds to the onverse. The same onsideration is valid or the B and Y olor pair. With the use o enters and surrounds at dierent sales, loated at oarser or

7 iner levels in the Gaussian pyramids, it is possible to ompute the olor opponenies at multiple sales. It is important to observe that in the previous visual salieny system the salienies were proportional to the dierenes between enter and surround regions, and here they are proportional to the ratios o these regions. Like Gevers and Smeulders [5] we use the logarithms o the spaes (R/G) and (Y/B), so we an ompute the opponenies by dierenes o logarithms aross the sales, instead o divisions. Additionally, as the logarithm o the inverse o an expression is a simple inversion o sign, we have only two pyramids or olor, one or ln(r/g) and the other or ln(y/b), saving the omputation o ln(g/r) and ln(b/y) pyramids. The proposed olor onstany/visual salieny algorithm, based on multisale olor ratio, onsists o the ollowing steps: 1. Conversion rom input RGB spae to opponent olor spae R G B Y, using equations (4) to (7). 2. Constrution o the ln(r /G ) and ln(y /B ) Gaussian pyramids, with 8 sale levels. 3. Computation o the multisale olor ratios with dierenes o logarithms at pyramid levels 2-5, 3-6, and Normalization o partial results through exponentiation (inverse o logarithm) and omposition o a resultant salieny map with the sum o all partial results. In the implementation o this algorithm the ln(r /G ) and ln(y /B ) are omputed as ln( R ) ln( G ) and ln( Y ) ln( B ). Indeterminations are avoided by taking ln(0)=0. The partial salieny maps obtained rom eah enter-surround sale (2-5, 3-6, and 4-7) and opponent olor ombination (R -G, G -R, Y -B, and B -Y ) are normalized through exponentiation beore ombining them to orm the salieny map. Sine this operation is the inverse o the logarithm applied to the soure signal, it restores the linear proportion between the partial maps. Experimental results obtained with the multisale olor ratio algorithm are shown in the ourth olumns o Figures 1, 2, and 3. In Figure 1, a better stability than using the ormer approahes is observed, although in the seond and third images remains some instability rom the greenish-yellowish trees and the yellowish gravel path. In Figure 2 the results are also more stable than the intensity normalization and omprehensive olor normalization ones, with the exeption o the irst image, where the yellowish bushes are not indiated as salient. In Figure 3 the results are also stable, although the reddish tree is not indiated as salient in the third image. This eet is due to the at that the salieny o the yellow lowers is approximately ten times higher than the salieny o the tree. This masking eet ould be avoided by onsidering separately the salieny maps o R -G and Y -B opponenies. 6 Perormane omparison As a benhmark evaluation, we applied the three tehniques to a soure RGB image o 512x384 pixels, using a PC omputer, with an AMD Athlon 800MHz proessor, 128Mb DRAM, under Windows 98, averaging 100 suessive exeutions o eah approah. Table 1 shows the exeution times obtained, where salieny detetion with our multisale olor ratio method presents lower exeution time than the other approahes. Table 2 shows the distribution o omputing time between the most important tasks arried out by the proposed algorithm. The omputation o entersurround dierenes is only 8% o the total exeution time, beause these dierenes are omputed at the sale o the enters, instead o at the soure image sale. For example, or a 512x384-pixel image, the entersurround dierenes between levels 3 and 7 are omputed using the dimensions o the enter image at the level 3 o the pyramids, i.e., 64x48 pixels. It an be observed that the task o initializing data strutures represents a signiiant omputational eort in the proessing o visual salienies. The better perormane o the multisale olor ratio algorithm ompared to the others is due to the ewer loat operations neessary to ompute the salieny, although the introdution o the logarithm operation is signiiant in the overall exeution time. Table 1: Exeution times or omputing visual salieny with the three dierent approahes. Approah Seonds Intensity normalization 0.86 Comprehensive olor normalization 1.19 Multisale olor ratio 0.77 Table 2: Distribution o exeution time between the tasks perormed within the multisale olor ratio approah. Task Fration o total exeution time Conversion RGB to R G B Y 0.21 Logarithm o R G B Y 0.22 Pyramids ln(r /G ), ln(y /B ) 0.23 Center-surround dierenes 0.08

8 Other tasks Conlusions In this paper we have ompared three approahes to olor onstany as applied to a landmark detetion system based on opponent-olor salieny. The irst approah, lighting intensity normalization through the transormation o olor rom RGB to hromatiity spae, has the drawbak o produing a loss o some inormation, sine it projets olors onto a lower dimensional spae. More speiially, sine eah o the normalized olors is linearly dependent, only two o the normalized olors (e.g., r and g) are enough to represent the rgb spae. Although this method is simple, it has shown an undesirable sensitivity to shadows and hanges in the illuminant olor. The omprehensive olor normalization has proven to be more stable to illumination hanges than the lighting intensity normalization, but presents higher omputational ost and also produes undesired hanges in the deteted salient regions. The olor onstany is aeted by the global olor measures in the image, and so the method is sensitive to the inlusion/exlusion o objets in the senes. We onlude that, or the target appliation, our method is more suitable than the other, beause it presents stable results with better orrespondene o salieny to the image regions that that we want to use as landmarks in outdoor environments. This harateristi is obtained with an eiient algorithm, whih presents a slight improvement in perormane ompared to the other approahes. Aknowledgments This work is partially supported by the Spanish Siene and Tehnology Commission (CICYT), in the sope o the projet Autonomous navigation o robots guided by visual targets, grant DPI C Reerenes [1] Berens, J. and Finlayson, G.D., Log-opponent hromatiity oding o olour spae, 15th International Conerene on Pattern Reognition, Barelona, Spain, 2000, vol. 1, pp [2] Berns, K.; Kepplin, V. and Dillmann, R. Terrain and obstale detetion or walking mahines using a stereo-amera head, 24th Annual Conerene o the IEEE Industrial Eletronis Soiety, 1998, IECON '98, pp [3] Brue, V., Green, P.R., and Georgeson, M.A., Visual Pereption, Psyhology Press, United Kingdom, [4] Finlayson, G.D., Shiele, B., and Crowley, J.L. Comprehensive olour image normalization, 5th European Conerene on Computer Vision, 1998, pp [5] Gevers, T. and Smeulders, A.W.M., "Color-based objet reognition", Pattern Reognition, 32(3), 1999, pp [6] Grossberg, S., Mingolla, E., Ross, W., "A neural theory on attentive visual searh: interations o boundary, surae, spatial, and objet representations", Psyhologial Review, 101(3), 1994, pp [7] Itti, L., Koh, C., and Niebur, E., "A model o salieny-based visual attention or rapid sene analysis", IEEE Transations on Pattern Analysis and Mahine Intelligene, 20(11), 1998, pp [8] Levitt, T.S. and Lawton, D.T., Qualitative Navigation or Mobile Robots, Artiiial Intelligene 44, 1990, pp [9] Murrieta-Cid, R., Briot, M., and Vandapel, N., Landmark identiiation and traking in natural environment, LAAS Report #98037, Toulouse, Frane, January [10] Otha, Y., Kanade, T., and Sakai, T., Color inormation or region segmentation, Computer Graphis and Image Proessing, 13, 1980, pp [11] Swain, M.J. and Ballard, D.H., Color indexing, Int. Journal o Computer Vision, 7(1), 1991, pp [12] Tan, T.S.C. and Kittler, J., Colour texture analysis using olour histogram, IEEE Proeedings o Vision Image and Signal Proessing, 141(6), 1994, pp [13] Todt, E. and Torras, C., Detetion o Natural Landmarks through Multisale Opponent Features, 15th International Conerene on Pattern Reognition, Barelona, Spain, 2000, vol. 3, pp [14] Wyszeki, G. and Stiles, W.S., Color Siene: Conepts and Methods, Quantitative Data and Formulae, John Wiley and Sons, U.S.A., [15] Institut de Robòtia i Inormàtia Industrial, CSIC-UPC, Barelona, Spain, The Argos robot, 2001, wwwiri.up.es/people/porta/robots/argos.