Visual Tracking of Athletes in Beach Volleyball Using a Single Camera

Transcription

1 Visual Tracking of Atletes in Beac Volleyball Using a Single Camera Tomas Mautner 1, Cristina Koc 2, Markus Tilp 2, Horst Biscof 1 1 Institute for Computer Grapics and Vision, Graz University of Tecnology, Austria 2 Institute for Sport Science, Karl-Franzens-University, Graz, Austria Abstract Tis paper aims at successful tracking of beac volleyball atletes during competition using only a single camera. Due to te wide range of possible motions and non-rigid sape canges, te tracking task becomes quite complex. We propose a novel metod based on integral istograms, to use a ig dimensional model for a particle filter witout drastic increase in runtime. We extend integral istograms to andle rotated objects. Additionally to te tracking process, a segmentation of te lower body parts enables generating real world player positions from a single camera view. Comparisons to and annotated position data revealed sufficient accuracy for classical sport scientific purposes. Te paper focuses on beac volleyball but te proposed metods can be utilized in oter sports and non sports applications. KEY WORDS: VISUAL TRACKING, BEACH VOLLEYBALL, TIME-MOTION ANALYSIS Introduction Wen analyzing sports games te main aspects of interest are te used tecniques, te played tactics and te pysiological demands of atletes. All tree caracteristics are important to quantify skills and sortcomings of atletes or teams and define requirements for training and competition. For te analysis of tecnique and tactics, video tecnology as become very common and is utilized in several ways to analyze tese aspects. Te simplest way is to use video recordings to provide feedback for atletes (Liebermann & Franks, 2004) or to study opponent teams during crucial game situations in replay. Beyond tese attempts interactive video systems (Dartfis, Fribourg, Switzerland or Statsot, Graz, Austria) are used to gater furter information e.g. by counting frequencies of special tecniques and by evaluating te effectiveness of actions. Suc an attempt was successfully used by Tilp, Koc, Stifter & Ruppert (2006) to generate video based statistics for te analysis and comparison of world class junior beac volleyball teams. Te positions of actions in sports are often crucial for success or defeat and terefore an essential information to rate te quality of an action. In order to determine te playing position it became quite common to define relevant zones of te court and to estimate in wic zone an action occurred (Huges & Franks, 2004). Getting accurate information about positions is complicated due to several factors. A distortion caused by te perspective view, missing marks and te transitions from one zone to anoter often causes wrong rating decisions and errors. 21

2 Exact position information is furtermore required to calculate covered distances, velocities and accelerations of atletes wit wic te pysiological demands can be estimated. Te main advantage compared to classical metods like eart rate monitoring or lactate testing is tat interaction wit te atletes can be avoided. Different metods for suc time-motion analyses ave been used since te early 1970 s. Before adequate tecnology was available suc analyses were made manually via observation (Reilly & Tomas, 1976) or via audio recording (Yamanaka, Haga, Sindo, Narita, Koseki, Matsuura & Eda, 1988). Due to accuracy reasons, metods based on video data followed by a manual computer supported analysis ave become te preferred metod in te last years (for review see Spencer, Bisop, Dawson & Goodman, 2005 or Bangsbo, Mor & Krustrup, 2006). Determining positions during interesting game situations, ratio of action and recovery time or te amount of pysically exausting actions like sprints or jumps would require an annotation of nearly eac frame of a video sequence. In order to obtain tis information wit a feasible amount of user interaction, an automatic system for position computation is needed. Figure 1. Te proposed metod can andle caracteristic player motions like jumps and digs, wic occur frequently during tracking. Commercial applications Existing commercial applications for more or less automated tracking and position estimation demonstrate te interest of teams and coaces in gatering suc information. Tey can be rougly divided into two groups: systems using markers (active or passive) and markerless systems. Two representatives for te first group are Cairos (Munic, Germany) and LPM (Abatec AG, Regau, Austria). Bot systems use a radio based metod, were every tracked object is equipped wit a transponder wic position is measured by several base stations. Tis tecnique as te advantages tat te 3D position is computed up to 1000 times per second wit a ig spatial accuracy and tat te number of tracked objects can be ig. A disadvantage of suc systems is te possible influence of markers on te atlete s beavior toug tis problem as improved remarkably by minimizing marker size. However, as most of te sport rules proibit te wearing of markers during competition te use of tese tecniques is very limited in sport practice. Markerless systems are mainly based on video input. Teir main advantages are tat te influence on players during competition is zero and tat also opponents can be observed. Furtermore, te obtained video data can also be used for feedback or tactical observations. One of te leading systems is Amisco Pro (Nice, France). It is a commercial multi camera matc analysis system (8 stable, syncronized and fixed camera orientations) approved by several European soccer clubs. Recently, te system as also been used scientifically to estimate covered distances and running velocities in international soccer as reported by Salvo, Baron, Tscan, Calderon Montero, Bacl & Pigozzi (2007). Altoug tis system may provide interesting data te required tecnical and financial effort (especially for ardware) is 22

3 an excluding factor for most type of sports. Terefore, it would be necessary to improve vision based tracking software to get accurate results witout an enormous amount of tecnical effort. Computer vision related work On te one and one can see tat computer vision and in particular tracking are increasingly important for digital game analysis. On te oter and many different games like soccer, ockey, tennis and oter type of sports ave been used as test data for new computer vision approaces. To andle te unpredictable beaviour of objects of interest during tracking sport games, e.g. atletes and ball, particle filter based metods ave become common in tat area. Since its introduction into te computer vision by Isard & Blake (1998), te particle filter as been used for various tasks and is a common metod for player tracking in sports. Te simplicity of te metod, te ability to recover from uncertainties during tracking, and te possibility of fusing different information cues in one tracker are major advantages of tis tracking metod (see Perez, Vermaak & Ganget, 2002; Perez, Vermaak & Blake, 2004). For te analysis of andball and basketball games Kristan, Perš, Perše & Kovačič (2006) ave developed an indoor tracking system. Due to ceiling mounted cameras, te mutual occlusions between players are minimized. Te tracking metod uses a color based particle filter, considering te player as an elliptical region. Te position of an object is ten estimated by te center of tracking ellipse. Okuma, Talegani, De Freitas, Little & Lowe (2003) combined a particle filter tracker wit te detection results of an offline trained classifier to track ockey players. Te tracking region of a player was defined to be an uprigt rectangle. Altoug te tracking results were quite impressive, results on estimated ground positions were not publised. A compreensive framework for automatic annotation of tennis matces was made by te group of Josep Kittler (Yan, Cristmas & Kittler, 2005). Te tracking of players was done by subtracting te current frame from a pre-computed background image and using a blob tracker on te results. In addition, a support vector macine is trained to detect tennis ball candidates and a particle filter is used to track te ball. Our approac Tis work presents our approac in beac volleyball for a vision based tracking system wic can be used in practice by trainers and atletes witout extensive tecnical effort. To acieve tese goals, te tracking algoritm sould be able to track atletes during competitions only by te use of a single camera and witout complex calibrations. Tis as te positive side effect tat already existing beac volleyball videos can be analyzed as well. Tracking information sould ten be used to compute real world coordinates wic provide exact position and enable time-motion analysis. We assume an offline annotation and tracking scenario, were te wole game video is available. Te tracking and position estimation process must not be fully autonomous. Terefore a small amount of user interaction for correction and re-initialization is acceptable. Due to te playing caracteristics of beac volleyball (and most oter game sports) and te constraints due to a single camera te tracking algoritms used ave to andle rotations and scale canges of bodies, e.g. squatting during a receive action. Specifically for beac volleyball te tracking metods sould provide position information to improve te rating of tecniques as well as motion analysis to estimate pysical load (e.g. by detecting jumping movements as seen in Fig. 1). Our approac consists of in tree main parts: configuration, tracking and position estimation. In te configuration step, te transformation between video image and court coordinates is 23

4 calibrated. Furtermore color and scale references are predefined for eac player, by marking tem in a single image. A background model is created automatically from input video. A color based tracker wic is computational efficient by using integral structures forms te second part. It allows rotations to follow players during all possible motions and estimates te size of a player using te data from te configuration step. Te tracker is only applied on te upper part of a player, wic stays more compact during motions. If player positions are needed an additional segmentation step, using a skin color classification wic was trained beforeand, is applied. Tis segmentation is only performed witin an area defined by te tracker, were te lower part of te body is assumed, and terefore te additional runtime is negligible. Real world coordinates are finally estimated using te calibration from te configuration step. Figure 2 visualizes te main parts and te work flow. Figure 2.Vizualization of te main processing steps. Te remainder of te paper is organized as follows. First, te particle filter approac is summarized and te transition model used for tracking is explained. Furtermore, te computation of te color properties and te likeliood function of te particles for te single object tracker are described. Based on te tracking results, players are segmented from te background to allow te computation of real world coordinates. Experiments sow an evaluation of tracking and position estimation results on manual annotated ground trut data. Finally, conclusion and summary are given at te end. Metods Tis section describes te tracking metod used in tis work. Te particle filter concept is briefly explained in addition wit te motion and appearance model used for player tracking. Object scale estimation and problems wit position estimation from single view cameras are illustrated. In order to obtain real world coordinates, an additional segmentation step is performed. Tracking wit particle filter Te idea of te particle filter is to estimate te state x t of a tracked object by using a set of weigted particles (Isard & Blake 1998). Eac particle simulates te beavior of te object using Monte-Carlo simulations, a motion model and a measurement. Given a state space model x t-1 at time t-1 and all measurements up to t-1 known as z 1:t-1 te posterior x t z 1:t ) can be estimated by te recursion of Equations 1 and 2 using te new measurement z t. 24

5 x t z1 : t 1) = Predict: xt xt 1) xt 1 z1: t 1 ) dxt 1 (1) z t xt ) xt z1: t 1 ) Update x t z1: t ) = (2) z z ) t 1: t 1 Te required posterior density function x t z 1:t ) of te new state can be approximated using sequential Monte Carlo simulations of a finite set of particles {x t i } i=1 Np. From an initial state, te weigts {w t i } i=1...np associated wit te particles are computed by sampling from a proposal distribution q(x t x t-1,z t ) (see Equation 3). w i t i i i i N zt xt ) xt xt 1) P were i i i q( x x, z ) t t 1 1: t i= 1 i w t = 1 (3) Using te state transition model x t x t-1 ) as proposal distribution leads to te bootstrap filter, were te weigts are directly proportional to te observation model z t x t ). Finally, N te posterior density can be approximated by P i i p ( x z ) t 1 : t i = wt x 1 t. To avoid te degeneracy of te particle set, resampling of te weigts is done if necessary (see Arulampalam, Maskell, Gordon & Clapp, 2002, for more details). State model used for players During te tracking process players are described wit rectangles given by center coordinates, size and rotation angle. In image coordinates te state model of an player at time t is defined by x t = [x t, у t, υx t, υy t, φ t ] were (x t,y t ) are te center coordinates of te rectangular window, (υx t, υy t ) are te velocities and φ t is te rotation angle of te player, see Figure 3. Te size of a player (,w) during tracking is computed directly, using te assumption of a fixed camera. Te Homograpy H between image coordinates and real world court coordinates is determined wit an initial manual calibration. If a player is annotated once as a reference, for example during initialization of te tracker, te scale parameters (,w in Figure 3) for different court positions can be estimated using te Homograpy H. Te real state of a player x t is estimated by a set of particles simulating possible states x i t. Applying an autoregressive model wit an constant velocity assumption, te transition probability x t x t-1 ) can be represented by: x = Ax + v t+1 t t (4) Wit tis model te motion of particles is defined by a drift component defined in matrix A, equal for all particles of a player, and a random component in v t, wic is assumed to be normally distributed for x, y and φ. Using te omograpy Assuming tat image coordinates are given for eac player, one would be interested in te real world coordinates. Reconstruction of full 3D coordinates of players or ball requires at least two cameras, wic we do not ave in our setup. However, under te assumption tat te players are moving on a specified plane one can use a perspective mapping between two planes, defined by te Homograpy, to compute court positions of players (for details see Hartley & Zisserman, 2002). 25

6 Figure 3. Left: Object description for players by a rectangular patc. For approximation of rotation te tracker is divided into tree sub-parts. Rigt: Image sows 3 possible states of particles wit different positions and orientations in blue, green, and red. x' = y' z' x y 1 (5) Estimating te unknown Homograpy matrix H in Equation 5, requires at least 4 points in te image and in te target view, respectively. Te linear transformation uses te omogenous coordinates of image points [x y 1] T to compute te transformation to te given target coordinates [x y z ] T. In te resulting rectified view perpendicular angles are reconstructed and, wit te metric world coordinates given, one can reconstruct real world court coordinates (e.g. 8x16m for beac volleyball, visualization in Figure 4). Figure 4. Left: Projection of real world points to te image plane of a camera and te transformation to a virtual top view projection. Rigt: Te known coordinates of te playfield corners are used for calibration of te setup. Screensot sows te perspective camera view and te undistorted top view image. Te differences of te resolution in dept are sown by te parallel lines in bot views. Wit te world coordinates in meters and given te frame rate of te video stream, approximations of speed and acceleration can be calculated. Te acievable accuracy of te field coordinates depends on te resolution of te camera as well as on te distance and orientation between player and camera. Players furter away from te camera center ave a lack of resolution, especially in te y-coordinates, and terefore less accurate positions can be calculated. 26

7 It as to be mentioned tat te Homograpy transformation contains only te transformation between planes, as sown in Figure 4. Terefore, it only olds for points on te calibrated playfield. By tracking players wo are not on te ground plane, te projected position would be wrong (see Figure 5). Figure 5. During jumps players are off te calibrated ground plane. Te estimated positions contain an error especially in teir y-coordinates, due to te assumed projection onto te ground plane. Color tracking To evaluate te set of particles, a measurement function as to be defined to see ow good a particle fits to te real state of a player. Color information is a simple but powerful metod to describe an object of interest. In contrast to sape description metods, wic ave also been used wit particle filters, color information is less vulnerable to clutter. In particular, te intensive and distinct team colors in sports support te use of color istograms for our model description. Using te HSV color space, an object is described wit 3 independent N B -bins istograms for te ue, saturation and value cannel. An object, in our case a player, is initialized wit tree reference istograms [ H ref, S ref, V ref] for te color cannels. To compare candidate istograms [ H P, S P, V P] sampled from a particle estimation wit te reference istograms, te Battacarrya similarity coefficient D( P, ref ) is used. Combining te color cannels te likeliood model z C x) is finally assumed as exponentially distributed wit a weigting constant λ as sown in Perez et al. (2002). Histogram creation for eac particle is a very time consuming task. Moreover, te particles overlap most of te time, so tat many image areas are described several times. Porikli (2005) computed te istogram information of an image using te integral image approac, wic leads to a drastic speed up. Additionally, te integral structure is only needed for te image area covered by particles, wic is usually muc smaller tan te wole image. Once te integral istogram is computed for an image, te istogram information of particles can be obtained using only tree operations independent from position and scale of te particle. Te disadvantage of using te integral structure is tat it cannot be rotated. Lienart & Maydt (2002), proposed a metod to compute 45 rotations in te integral image wic is not sufficient for our aims. Barczak, Jonson & Messom (2006) extended te set of possible rotations to any angle by approximating from pre-computed rotated images. Applying suc an approac to a uge set of particles wit different rotations would diminis te speed up acieved by te integral approac. We decided to use an approximation approac similar to Grabner et al. (2006). Te original tracking rectangle is divided into N S subparts to approximate te rotation in te integral image (see Figure 3). Assuming tat te subparts are independent, te color likeliood for a particle wit state x and consisting of N S subparts is finally computed by: 27

8 N S C 2 C C z x) = ex λ D (, )) (6) { H, S V} j= 1 C, Te improper approximation of te object due to te rotated subparts is compensated by te ig number of available particles. In addition, a spatial relation is integrated into te likeliood computation of te particles, wic was also sown by Perez et al. (2002). Tis leads to more stable tracking results. Furtermore, te number of subparts and teir spatial relation can be canged. Including information about background Usually, kernel or mask functions are applied to take into account tat some background pixels are always included in te tracking window. To measure te influence of background pixels in our integral approac, a background probability z B x) is included in te formulation of te measurement likeliood of te particles. Because of te static camera, te background image can be computed in a preprocessing step. Using Equation 6 also for te background similarity, te final observation model for a particle is given by: z x) z C C z x ) x ) + z B j, P x) j, ref = (7) For every particle i wit state x i t at time-step t, te background similarity D( i j,p, i j,b) is measured for eac subpart j. Te istogram i j,p is sampled from te actual frame and i j,b is computed for te same area in te background image. Te integral structure for te background as to be computed only once beforeand (see Figure 6). Integrating te background probability prevents te tracker from drifting into background regions during mutual occlusions of te players. Figure 6. Top-row: Left: Actual input frame, were te green rectangle marks te patc used for te reference color istogram. Middle: Probabilities of pixels being te tracked object, using only color information. Similar colored objects in te background cause errors. Rigt: Pre-calculated background image. Bottom-row: Left: Probability of areas to be background. Rigt: Combined probabilities were background areas wit colors similar to te tracked players, like advertisement spaces, are less likely now. 28

9 Segmentation and position estimation Based on te obtained tracking results wic are valid for te upper part of te atletes, te computation of te real field position is bounded on a smaller area. Knowing te position of te player, te rotation-angle of its body and an estimated scale, an area for segmentation is defined in te video frame. Te similarity between skin colored and sand colored pixels makes it ard to segment te patc in player and background. We cose to transform te RGB pixel into te YCbCr color space, wic as sown good performance for face or skin segmentation tasks (Pung, Bouzerdoum & Cai, 2005). Using about 2 millions of skin, sand and background training pixels, a mixture of Gaussian models ave been computed in an offline process to describe te different classes in te YCbCr color space (Figure 7 sows some segmentation results). Figure 7. Left: Original input video frame. Middle: Segmented sand-colored pixel. Rigt: Skin segmentation can be used to create more accurate information about player positions. Using te segmentation results of skin colored pixels and pixels containing to te player, known from tracking, te image can now be divided into player and background regions. Morpological operations are used to filter out small segmentation errors. Te final segmentation result, wic can be seen as an example in Figure 8, is a combination of te biggest segmented regions. Te estimated ground position in image coordinates (x,y) of a player is computed by te mean of all x-coordinates of te segmentation and te maximum y-coordinate. Tis is motivated by te fact tat te mean represents te center of gravity of te region, respectively te player. Te maximum y-coordinate is taken because of te used projection from image coordinates to real world coordinates. One can see tat te combination of tracking and segmentation results leads to more accurate results. Assuming a fixed size for te lower part of te player, or using only a rectangular window, would only be valid for players standing uprigt. Evaluation experiments and first results Te following section contains an evaluation of te proposed tracker. Te metod is compared to manual annotations in terms of overlaps on players and estimated field positions. Terefore, a set of 12 test-sequences, eac consisting of several undred frames was used. All sequences, including male and female rallies, ave been annotated manually at every tird frame. Additional results for multi-object tracking in sport applications can be seen in Mautner & Biscof (2007). 29

10 Figure 8. Segmentation of patces into player and background regions. Size and position of te segmented region are derived from te tracker result. Estimated positions are indicated by arrows. Verification of tracker results Te reference ground-trut was manually annotated by experts familiar wit beac volleyball. A rotated rectangle was placed over te upper part of te player by te annotators in every tird frame of te test videos. An overlap factor is computed from te sared area between te manual reference and te tracking result in relation to te total area of bot rectangles (Figures 9 and 10). Total overlap of reference annotation and tracking result in an overlap factor of 1 and no overlap leads to a factor of 0. Figure 9. Left: Progression of te overlap factor during tracking of two players during sequence 1. Rigt: Manual annotation (blue rectangle) and tracking result (red dased rectangle) for frame 346 in sequence 1. Bot players are fully covered by teir trackers, but according to te bending of player 1 te overlap factor is about 0.4, and lower tan for player 2. As described in te metod section, te result of te tracker is computed over a weigted sum of te particles. Te size of eac particle is estimated from its position, and terefore, te size of te tracker result is always a combination of different scales. Additionally, te size of a 30

11 torso is assumed to be fixed, wic is not true, in image coordinates, if te players bend or crouc during te game. Tese mentioned difficulties do not exist if annotation is performed by a uman, and terefore experts always scaled teir reference rectangles to te visible part of te upper body. Low overlap values during suc frames can be traced back on te differences between uman perception and automatic computation (see Figure 9 and Figure 10). Figure 10. Left: Tracker overlaps on all test sequences, given by mean values and standard deviations. Te sequences 2-6 belong to female games for wic automatic tracking is less accurate due to te smaller amount of specific colors and te similarity between sand and skin color. Rigt: Manual annotation and tracking result for frame 133 of sequence 1. A notable difference between sequences extracted from male and female games can be seen in Figure 10. Sequences from 2 to 6 are taken from female games, wile sequence 1 and sequences 7 to 12 are from male competitions. Tis effect can be explained by te different appearances of te atletes. Male atletes wear sirts wic, most of te time, are distinguisable from te background due to specific team colors. Because of te similarity between skin and sand color, te appearance of female atletes is not tat precise during tracking wic makes it arder to estimate te correct scale of a player. Neverteless, te overall performance is balanced over te sequences. Considering te example frames given for sequence 1, an overlap factor of 0.45 still means an acceptable tracking result. Please note tat no manual interaction was needed during te 12 test scenes. Estimation of field positions Similar to te tracker evaluation te reference position data consists of manual annotations. For every tird frame of te test sequences, an image coordinate per player was defined as te reference position. Te difference between manual reference and automatic position result is given as te Euclidian distance in image and real world coordinates. Results of estimated positions are directly compared wit te manual ground trut, witout any additional filtering. Te variance in te results could be reduced, if situations like jumps and occlusions between players would be excluded or corrected manually. Note tat even varying uman annotations can result in different position results. Figure 11 sows a visual comparison between estimated positions and reference annotations. Trajectories of one team during a rally are sow individually in two separate top-view projections. 31

12 Figure 11. Comparision between manual annotation and results of our metod. Positions are sown for every tird frame only. Left and rigt images sow te manually in comparison wit automatically generated trajectories of two different players during a rally. Note left image: Projection causes a wrong position estimation into te opponent field during a jump movement (left side of te field). After te landing te player position was estimated correctly again. As previously observed, several problems may occur wen computing an exact atlete position from a single camera view. Depending on te point wic is projected to real world field coordinates, te position can vary. Even te distance between bot legs can be around alf a meter. Additionally, te accuracy of te projection depends on te geometric resolution per pixel. In te used test sequences, te resolution varies between 3 cm/pixel and 10 cm/pixel, depending on te distance to te camera. Considering tis fact, te results sown in Figure 12 are satisfying. Compared to te tracking results no difference between male and female games is observed. Tis effect can be led back to our special skin segmentation step. Neverteless, te results are accurate enoug to answer several sport scientific questions and can be used for furter analysis. Based on tis position data, oter parameters suc as velocity and acceleration can be derived. Using te projected coordinates of te tracker, te resulting speed during jumps is not realistic due to projection errors (see Figure 11). Furtermore, a caracteristic motion occurs, consisting of acceleration away from te camera followed by te inverse motion back towards te camera. Te wole jump motion takes place in a maximum timeslot of about 30 frames. Suc a saped pattern can easily be found in te provided velocity data of eac player and terefore could be exploited to detect jumps. Conclusion and furter work A simple and yet effective metod as been presented for tracking multiple objects witin te scope of sport applications. Te presented approac aims at obtaining position and motion information using te video input of a single camera, as tis is te typical situation in sports practice. Tis aim could be acieved by combining several computer vision metods. By dividing te tracker window into subparts, te approximation of rotations in te integral istogram is possible. Terefore, sport specific motions can be followed wit almost no additional runtime compared to using only an uprigt rectangle. Tracking results and segmentation of skin colored regions are combined to estimate real world court positions of 32

13 atletes. Togeter wit te possibilities of using te calibration for scale estimation during tracking and computation of real world coordinates, we are able to create useful tracking and position results for beac volleyball games. Figure 12. Comparison of manual annotated positions and te results of te automatic metod in image and real-world coordinates. Te main advantages compared to existing metods are te rotation sensitivity, wic delivers tracking results more similar to uman annotation, and te combination of tracking and segmentation. As sown in te results, our metods deliver more information to analyze player positions tan common rectangular or elliptical saped trackers. We believe tat te presented metods, togeter wit a reasonable amount of manual interaction, are sufficient for motion analysis and te evaluation of pysical demands in beac volleyball. Preliminary results indicate tat it will be possible to detect frequency of jumping movements automatically in future. Furtermore, position data can be used to improve accuracy of action annotations and tactical analyses. Te presented results are valid for beac volleyball but can be principally transferred to similar outdoor sports were fixed multiple camera systems are not available. A successful application of te presented metod would be a great relief for te annotation process in several types of sports. Te integration of te proposed metods into existing game analysis software is te next step. Tracking and position data sould be combined wit expert annotations about player beaviour and used tecniques. Acknowledgement Tis work as been supported by te FWF (Austrian Science Fund, P18600). References Arulampalam, S., Maskell, S., Gordon, N., & Clapp, T. (2002). A Tutorial on Particle Filters for Online Nonlinear/Non-Gaussian Bayesian Tracking. IEEE Transaction on Signal Processing, 50(2), Bangsbo, J., Mor, M., & Krustrup, P. (2006). Pysical and Metabolic Demands of Training and Matc-Play in te Elite Football Player. Journal of Sports Science, 24(7), Barczack, A., Jonson, M., & Messom, C. (2006). Real-time Computation of aar-like Features at Generic Angles for Detection Algoritms. Researc Letters in te Information and Matematical Science, 9,

14 Grabner, M., Grabner, H., & Biscof, H. (2006), Fast approximated SIFT. Proceedings of Asian Conference on Computer Vision, Hartley, R., & Zisserman, A. (Eds.) (2000). Multiple View Geometry in Computer Vision. Cambridge University Press. Huges, M. & Franks, I.M. (2004). How to develop a notation system. In Notational Analysis of Sport, 2nd Ed., (edited by M. Huges, & I.M. Franks), New York, Routledge. Isard, M., & Blake, A. (1998). Condensation Conditional Density Propagation for Visual Tracking. International Journal of Computer Vision. 29(1), 5-28 Kristan, M., Perš, J., Perše, M., & Kovačič, S. (2006). Towards Fast and Efficient Metods for Tracking Players in Sports. CVBASE 06 Proceedings of ECCV Worksop on Computer Vision Based Analysis in Sport Environment, Liebermann, D. G., & Franks, I. M. (2004) Te use of Feedback-based Tecnologies. In: Huges, M., Franks, I. M. Notational analysis of sport. Routledge, London/New York Lienart, R., & Maydt, J. (2002). An Extended set of aar-like Features for Object Detection. Proceedings International Conference on Image Processing, Mautner, T., & Biscof, H. (2007). A Robust Multiple Object Tracking for Sport Applications. Proceedings of Austrian Association for Pattern Recognition, Okuma, K., Talegani, A., De Freitas, N., Little, J., & Lowe, D. (2004). A Boosted Particle Filter: Multitarget detection and tracking. Proceedings European Conference on Computer Vision, Perez, P., Vermaak, J., & Ganget, M. (2002). Color-based Probabilistic Tracking. Proceedings European Conference on Computer Vision, Perez, P., Vermaak, J., & Blake, A. (2004). Data Fusion for Visual Tracking wit Particles. Proceedings of IEEE (issue on State Estimation), 92, Pung, S., Bouzerdoum, A., & Cai, D. (2005) Skin Segmentation using Color Pixel Classification: analysis and comparison. IEEE Transaction on Pattern Analysis and Macine Intelligence. 27(1), Porikli, F. (2005). Integral Histograms: A Fast Way to Extract Histograms in Cartesian Spaces. Proceedings IEEE Conference on Computer Vision and Pattern Recognition, 1, Reilly, T., & Tomas, V. (1976) A Motion Analysis of Work-rate in Different Positional Roles in Professional Football Matc-play. Journal of Human Movement Studies, 8, Salvo, Di, V., Baron, R., Tscan, H., Calderon Montero, F.J., Bacl, N., & Pigozzi, F. (2007) Performance CaracteristicsAccording to Playing Position in Elite Soccer. International Journal of Sports Medicine, 28, Spencer, M., Bisop, D., Dawson, B., & Goodman, C. (2005). Pysiological and Metabolic Response of Repeated-spring Activities Specific for Field-based Team Sports. Sports Medicine, 35(12), Tilp, M., Koc, C., Stifter, S., & Ruppert, G. S. (2006). Digital Game Analysis in Beac Volleyball. International Journal of Performance Analysis in Sport, 6(1), Yamanaka, K., Haga, S., Sindo, M., Narita, J., Koseki, S., Matsuura, Y. & Eda, M. (1988). Time and Motion Analysis in Top Class Soccer Games. In T. Reilly, A. Lees, K. Davids, W. J. Murpy (Eds.). Science and football, , London: Spon. Yan, F., Cristmas, W. & Kittler, J. (2005). A Tennis Ball Tracking Algoritm for Automatic Annotation of Tennis Matc. Proceedings of te Britis Macine Vision Conference, 2,