Page  00000423 Real-time analysis of expressive cues in human movement 1 Antonio Camurri, Riccardo Trocca, Gualtiero Volpe Lab InfoMus - DIST, University of Genova, http://infomus.dist.unige.it Abstract. This paper focuses on our recent research on (i) the real-time analysis of expressive cues in human movement from visual information, (ii) the reconstruction of emotional, KANSEI higher-level information, (iii) the mapping of recognised expressive and emotional cues on different outputs (eg visual, audio), in the framework of novel mixed-media applications. This research is part of the EU IST project MEGA (Multisensory Expressive Gesture Applications, www.megaproject.org) facing realtime analysis of expressive gesture in artistic contexts (with particular reference to non-verbal communication through human movement and music signals), and multimodal/cross-modal mapping of expressive content. An ultimate goal of our research is to contribute toward a deeper understanding of the relations between gesture, music, and visual languages, and toward novel conceptual models and practical tools for participative, shared interactive performance spaces. Emotional and expressive content plays a fundamental role in this scenario. In a more general perspective, research aims at enhancing man-machine communication in mixed reality, mixedmedia and collaborative environments by adding a new channel: expressiveness. The analysis methods and computational models are inspired by several sources, including humanistic theories of non-verbal communication, such as Rudolf Laban's Theory of Effort, developed for dance and choreography and Shaeffer's Morphology developed for music ("musique concrete"). Consolidated research results are transferred into software libraries for our Eyes Web open software platform freely available from the web site www.eyesweb.org. 1. INTRODUCTION KANSEI Information Processing has been proposed as the third target of information processing. In an influential paper (Hashimoto, 1997), it is stated that the first target of information processing is the physical signal, i.e. sound, light, force, the second is language, the field of logic, of symbolic knowledge, the third is KANSEI, that refers to feelings, intuition, sympathy. KANSEI Information Processing deals with information that is related, but does not belong, to the domain of physics and logic. When a message in human communication is transmitted, three channels are used: physical, logical (semantic, the explicit content) and KANSEI, that carries information about the sender, his personality and intentions, or, in other words, his KANSEI. The word KANSEI refers to characteristics of the message that can evoke an emotional response, or a feeling, that are interconnected with perception. In this perspective, the development of tools dealing with KANSEI would greatly improve the actual state of human-computer interaction that suffers several deficiencies in the expressive field. One of the main goals of our research described in this paper is to face, from a scientific viewpoint, KANSEI communication in artistic performance. The target is to discover methods suitable to build a system that can evaluate KANSEI in human gesture. Gestures carry what Cowie (2001) calls implicit messages: the same action can be performed in several different ways, stressing different qualities of the movement, expressing feelings, moods, intentions. It is also possible to recognize a person from the way she moves. The attention is focused in this paper on interactive dance, since it is an artistic expression of human gesture with strong potential of emotional communication and arousal on spectators. 2. METHODOLOGY The ultimate goal is to build a system able to recognise and analyse KANSEI features of a certain phenomenon, like music or dance performance. A set of evoked emotions, or the arousal on a spectator can be seen as subset of KANSEI and described through the use of an emotional space. In this perspective, the model of such an evaluation system consists of the following components: 1. KANSEI function: it maps features of the studied phenomenon in an emotional space. This function should model the interactions between the physical world and the emotional space, emulating the effects that certain physical features of the studied event would have on the evoked emotional response. 2. Translation Function of a point in the emotional space. For example, a function expressing the distance of a point from a set of known/important emotions in that space (e.g. in the well-known circumplex model valence/arousal). The definition of the KANSEI or emotional space, and the labelling of relevant points, e.g. in terms of basic emotions, is a crucial issue. Another issue concerns the modeling of the KANSEI function, probably the most difficult problem in KANSEI Information Processing. This modeling can be based on different approaches: for example, neural networks or clustering algorithms. In this way, the power of neural networks is used to find non-linear relations between physical measures and KANSEI space. In fact, it is thought that KANSEI space might be nonlinear and this fact can justify the use of such non-linear techniques. An example may be found in the work of Suzuki (Suzuki, 1997) focused on sound perception, where a neural network is trained to place its output in a sort of KANSEI space. Another possible approach would involve the creation of an explicit description of the studied phenomenon. For example, starting from a human movement signal, to reconstruct a description in terms of expressive cues (such as fluentness, directness, energy, etc.), shapes, and phrasing. As an analogy, in music this would be equivalent to recreate a "score" starting from a sound signal, or, better, build a representation of the signal in terms of a vocabulary similar to Shaeffer's morphology (1977). Such symbolic description would allow a system to detect perceptual patterns in the studied event (or series of events) and capture more complex structural relations that might require the use of logics. 1 This research is partially funded by the EU IST Projects CARE HERE (Creating Aesthetically Resonant Environments for the Handicapped, Elderly and Rehabilitation) no. IST-2001-32729, and MEGA (Multisensory Expressive Gesture Applications) no. IST-1999-20410 (www.megaproject.org), and by National Projects COFIN2000 and CNR project CNRG0024AF "Metodi di analisi dell'espressivita nel movimento umano per applicazioni in Virtual Environment". 423

Page  00000424 The approach we adopted finds its foundations on several different sources. Main contributes come from: 1. Research and theories on KANSEI and emotion arousal/appraisal (e.g., the Hashimoto's theory on KANSEI Information Processing sketched above); 2. Biomechanics and motion capture techniques; 3. Research and theories from artists on communication of expressiveness in dance (e.g., Rudolf Laban's Theory of Effort, Laban 1947) and music (e.g., Pierre Schaeffer's Morphology, Shaeffer 1977); 4. Research and theories from psychology on non-verbal communication of expressiveness (Wallbot 1980, Argyle 1980); 2.1 Conceptual Architecture This research is grounded on the concept of gesture, conceived as the vehicle carrying a set of temporal/spatial characteristics that are responsible of conveying expressiveness. In this perspective, this work adopts the general guidelines of the layered conceptual framework for expressive gesture applications described in (Camurri, De Poli, Leman, Volpe, 2001), and discusses the relations to the previously mentioned KANSEI approach, on the special case of human movement analysis. In the case of a movement performance, the performance is thus divided in a sequence of gestures where gesture's boundaries are detected by studying velocity and direction variations. Analysis is performed through different layers/steps following a bottom-up approach: Layer 1 - Physical Signals: This is the information that is captured by the sensors of a computer system. Physical signals may have different formats strongly dependent on the kind of sensors used to study movement.. For example, they may consist of sampled signals from tactile, infra-red sensors, signals from haptic devices, or low-level data frames in video. In this context the word "sensors" is related to the physical sensors employed and to the algorithm used to extract a given set of low level data. We can therefore speak of "virtual sensors" or "emulated sensors". A CCD camera can be an example of a physical sensor, while the optical flow or the motion templates or the positions of certain points in the frame sequence are examples of data extracted from "virtual sensors" implemented by the cited algorithms. Layer 2 - Low-level features and statistical parameters: Measures for a collection of motion cues describing the movements being performed are calculated. The extracted values of the motion cues can be processed by means of statistical methods. Examples of these low-level features are the amounts of contraction/expansion, stability, rotational movements... A particularly important set of cues are the ones related to the effort dimensions described in Laban's Theory of Effort (space, time, weight, and flow): for example, the amounts of straightness (i.e., how much a movement is direct or flexible), impulsiveness (with respect to which a movement can be quick or sustained), fluency (bounded or unbounded movements). Layer 3 - Mid-level features and maps: "In this layer, the purpose is to represent expression in gestures by modelling the low-level features in such a way that they give an account of expressiveness in terms of events, shapes, patterns or as trajectories in spaces or maps." (Camurri et al., 2001). Data from several different physical and virtual sensors are likely to be integrated in order to perform such a step. A movement sequence is divided in gestures. Each of them is characterized by the measures of the different cues extracted in the previous step (e.g., speed, impulsiveness, straightness, etc.) The problem here is to identify those strokes in a complex sequence and associate to them the qualities deemed important for expressive communication. The output can be a sequence of trajectories in a semantic space (i.e., a gesture is seen as a trajectory in a map representing categories of semantic features related to emotion and expression on a pre-defined grid; a sequence of gestures is associated to a sequence of trajectories in the map). Another possible output is a symbolic description of observed strokes and measurements of several quantities describing them. Layer 4 - Concepts and structures: for example, emotional content and KANSEI concepts (e.g., basic emotions fear, grief... information on arousal; intentional gestures such as Laban's types of effort "pushing", "gliding",...). This high-level information is built from low-level and mid-level features, using various analysis techniques (statistical, time series, etc). Following the scheme depicted in the previous section, the KANSEI function lies in the first three layers, while the Interpretation Function is a main concern of Layer 4. 2.2 Microdances A reference archive of microdances has been created and studied. With "microdance" it is meant a short video fragment containing enough information to be able to recognize expressive features. Human testers evaluate each microdance. Microdances are used to evaluate and test the developed algorithm by comparing the outputs of the computational models with spectators' rating of the same dance fragment. Microdances can also be useful to isolate further factors related to KANSEI and expressiveness or to provide an experimental evidence with respect to the cues that choreographs and psychologists already identified: this is mainly obtained by an analysis of differences and invariants in the same choreography performed with different expressive intentions (for example, a comparison can be done between a choreography performed in a "neutral" way, i.e., didactically and without any expressive intention, and the same choreography performed with expressive intentions corresponding to the four basic emotions fear, grief, anger, and happiness). 2.3 Subtractive analysis approach One of the main challenges is to discover techniques and identify basic factors of KANSEI and deep emotional arousal. To this aim, an approach we are working on is based on the live observation of genuinely artistic performances, and their corresponding video recordings. A reference archive of artistic performances has to be carefully defined for this method, chosen after a strict interaction with composers and performers. Image processing techniques are utilized to gradually subtract information from the video recordings. For example, parts of the dancer's body could be progressively hidden until only a set of moving points remain, deforming filters could be applied (e.g., blur), the frame rate could be slowed down, etc. Each time information is reduced, spectators are asked to rate the intensity of their "arousal" in a scale ranging from negative to positive values (a negative value meaning that the video fragment would rise some feeling in the spectator but such feeling is a negative one). The transitions between positive and negatives rates and a rate of zero (i.e. no expressiveness was found by the spectator in the analyzed video sequence) would help to identify what are the movement features carrying expressive information. A deep interaction is needed between the image processing phase (i.e. the decisions on what information has to be subtracted) and the rating phase. This subtractive approach is different from the previous studies by Johansson (1973) and more recently by Cowie (2001), where it is demonstrated that a limited number of visible points on human joints allow an observer to recognise information of movement. Our subtractive method is currently subject to investigations and experiments. The important fact is that the feedback from such experiments provide information on which movement cues our research should further investigate. The cues described in the 424

Page  00000425 following sections are also motivated by the results of these preliminary experiments. 2.4 Analysis Perspectives: Kinesphere and General Space According to the choreographer Rudolf Laban, a main distinction has been made between the analysis of movement in the Personal Space, referred also as Kinesphere, and the analysis of movement in the General Space. In "Modem Educational Dance" (Laban 1963, p. 85) Laban wrote: "Whenever the body moves or stands, it is surrounded by space. Around the body is the sphere of movement, or Kinesphere, the circumference of which can be reached by normally extended limbs without changing one's stance, that is, the place of support. The imaginary inner wall of this sphere can be touched by hands and feet, and all points of it can be reached. Outside this immediate sphere lies the wider or "general" space which man can enter only by moving away from their original stance. He has to step outside the borders of his immediate sphere and create a new one from the new stance, or, in other words, he transfers what might be called his "personal" sphere to another place in the general space. Thus, in actual fact, he never goes outside his personal sphere of movement, but carries it around with him like a shell." Movement is therefore considered under two different points of view: 1. Detailed movement of a single person (e.g., the movement of the centre of gravity or joints of a dancer) in his own "Kinesphere" or "Personal Space"; 2. Movement of one or more persons in a wider space, the "General Space" (e.g., a group of dancers moving on a stage, a group of visitors in a museum exhibit). 3. EXAMPLES OF ANALYSIS IN THE PERSONAL SPACE The methodology sketched above has been applied to analyse movement of dancers both in the Personal Space and in the General Space. Here some examples of analyses in the Personal Space are presented organized on different levels according to the described bottom-up approach: (i) Processing of low-level data coming from a camera (Layer 1): background subtraction techniques are used in order to extract the dancer's silhouette. The resulting images are then used to calculate Silhouette Motion Images (SMI). (ii) Extraction of low level features and parameters (Layer 2): in particular, as examples of possible cues, the quantity of motion and the contraction index will be presented and discussed. (iii) Segmentation of movement in motion and pause phases (Layer 3) by using the quantity of motion calculated in the previous layer. Examples of gesture representations by means of suitable feature spaces and/or symbolical descriptions will be given later in this paper. Analysis is performed in real-time on a set of recorded microdances (but it can be easily repeated on live dance performances) using a collection of software modules implemented in the framework of the EyesWeb open architecture for expressive gesture processing (Camurri et al, 2000). 3.1. Layer 1: Silhouette Motion Images (SMI) A Silhouette Motion Image is an image carrying information about variations of the silhouette shape and position in the last few frames. SMIs can be seen as a special case of Motion Templates (see the OpenCV Reference Manual, available at http://www.intel.com/research/mrl/research/opencv/), where information about time is implicit in the image and not explicitly recorded. The SMI is generated by the following formula: motion _image[t] = silhouette[t - i]- silhouette[t] The motion image at frame t is generated adding together images of the silhouette in the previous n frames and then subtracting the silhouette at frame t. The resulting image contains just variations happened in the previous frames. If n is the number of frames in which the SMI is calculated and n=l, then the SMI carries information about the instantaneous variations of the silhouette. Working with a higher n allows capturing more information about the shape of motion and results are smoother, because the effect is similar to filtering. Figure 1 shows a SMI with n=4. In the figure the SMI is the gray area, while the darker contour shows the latest silhouette in order to make the figure more understandable. Figure 1. SMI with n=4. 3.2. Layer 2: Quantity of Motion The simplest use of a SMI is calculating its area. The result can be thought as a rough approximation of the quantity of motion, i.e. q=m * v, where m is the mass and v stands for velocity. Of course the area of a SMI is not q, but the behavior is similar: actually the shape of the graph is close to the shape of the graphs of velocity of a marker put on a limb. However the SMI area alone is not a very reliable measure of movement, first because it suffers the same limitations of silhouette ("internal" motion is not detected), second, it is strongly dependent on the dancer's distance from the camera, third it is difficult to compare results of different dancers. It is possible to partially overcome the last problem scaling the SMI area by the area of the most recent silhouette. Movemen t=Area(SMI[t, n])/Area(Silhouette[t]) In this way the measure becomes almost independent from the camera's distance and it is expressed in terms of fractions of the body area that moved. For example it is possible to say that at instant t a movement corresponding to the 2.5% of the total area covered by the silhouette happened. 3.3. Layer 2: Contraction Index The contraction index is a measure, ranging from 0 to 1, of how the dancer's body uses the space surrounding it. We define a bounding box that surrounds the dancer's whole body and compare the area covered by this rectangle with the area actually covered by the silhouette. Intuitively, if the limbs are fully stretched and not lying along the body, the contraction index will be low, while, if the limbs are kept tightly nearby the body, the contraction index will be high, near to 1. While the dancer is moving, the contraction index will vary continuously. An interesting feature of the contraction index is that if it is used with data from just one camera, its information are still reliable, being almost independent from the distance of the dancer from the camera. Of course, in case of too long distance, image quantization problems appear. 425

Page  00000426 Figure 2. Silhouettes and their bounding boxes The leftmost one has high contraction index while the other has low contraction index. Figure 2 shows two examples of silhouette, displayed with their bounding boxes, with high and low, respectively, contraction indexes. A possible use of this parameter consists of sampling its values at the end and beginning of a stretch of movement, in order to classify that movement as a contraction or expansion. 3.4. Layer 3: Motion segmentation and gesture representation The silhouette motion image has interesting properties: the evolution in time of its (normalized) area (what we called quantity of motion) resembles the evolution of velocity of biological motion, which can be roughly described as a sequence of bell-shaped curves (motion bells). In order to segment motion by identifying the component gestures it is interesting to extract a list of these motion bells and their characteristics, i.e. peak value and duration. This can be also useful to obtain a first symbolical description of motion. One of the problems with the silhouette motion image approach, even dividing it in two vertical halves, is that several different movements may result super-imposed to each other, resulting in several motion bells to be overlapped. It is necessary to separate those motion bells in order to have a better description of motion. A first attempt at this consists in recognizing phases during which the dancer is moving (motion phases) and phases during which the dancer does not appear (i.e, movement is not perceived by a spectator) to move (pause phases). Actually, even if the dancer seems not to be moving, very small movements occur and they are detected by the motion image (together with some noise). After some attempt an empirical threshold has been defined, and the dancer is considered to be moving if the area of the motion image is greater than 2.5% of the total area of the silhouette. Figure 3 shows motion bells after segmentation: a motion bell characterizes each motion phase. Figure 3. Motion segmentation Bells may also be represented as trajectories in suitable spaces. For example, figures 4a and 4b show a running EyesWeb patch in which features are represented as trajectories in a 2D space. The dimensions are quantity of motion and fluentness. L -,:)................. = _: i ^'.., "" 'j^.. i wd (b) Figure 4a and 4b. A running EyesWeb patch: gestures are represented as trajectories in a 2D space. In figure 4a the dancer is not moving: the current position in the space (window in the right) is moving toward the bottom left parts of the 2D space (yellow stripe), a position characterized by low quantity of motion and low fluentness (i.e., the amount of pause phases is dominating the amount of motion phases). In figure 4b, a high-energy gesture is displayed. The red shadow around the dancer (the SMI) in the upper-left window of figure 4b is proportional to the quantity of motion and the position in the space (the yellow stripe in the right window) is moving toward the top-right region in that window, characterized by high quantity of motion and high fluentness. It is interesting to notice that the motion bells approach is currently applied also to sound signal analysis to investigate interesting cues candidate for expressive/KANSEI analysis. 3.5. Toward a symbolic description of human movement Another output of Layer 3 is a symbolic description of the movement, which can be further analyzed to produce inferences on the underlying emotional content (e.g., in terms of basic emotions expressed by the dancer). The following example is obtained by using the "quantity of motion" to operate segmentation and the "contraction index" to distinguish between contraction and expansion phases of movement: pause(9, 32). expansion( 41, 4, 0.032, 0.023, 2, 0.034, -0.087, 0.40). pause(45, 6). contraction(51, 12, 0.035, 0.035, 9, 0.080, 0.092, 0.15). pause(63, 64). 426

Page  00000427 contraction(127, 5, 0.036, 0.024, 2, 0.039, 0.76, 0.83). The meaning of each term in the symbolic description above is the following: * Pause(startJrame, length) corresponds to a phase of stillness that started at startJrame and lasted length frames. * Contraction(startJrame, length,start_value, end_value,peak_v alue_offset,peak_value, contraction_index_delta, start_contrac tion_index) corresponds to a phase of motion, started at frame start Jrame and lasted length frames. First and last SMI area values are start value and end_value, while peak_value_offset says after how many frames the peak value peak_value was found. Finally contraction_index_delta shows the variation of the contraction index between the start and the end of the phase, and start_contraction_index says what was the value of the contraction index at the first frame of this phase. * Expansion(...) the same has before, but while contraction(...) has a contraction_index_delta greater than 0, for an expansion it is less than 0. 4. IMPLEMENTATION: THE EYESWEB MOTION ANALYSIS LIBRARY the work described in the previous sections resulted in the design and implementation of a collection of software modules for the EyesWeb open architecture for expressive gesture processing (Camurri et al, 2000; see also www.eyesweb.org). These and other modules are grouped as a separate library of EyesWeb blocks and patches for real-time analysis of expressive cues: the Eyes Web Motion Analysis Library. It includes: (i) Blocks and patches for extraction and pre-processing of physical signals (typically, video frames from cameras): e.g., feature tracking using the Lucas and Kanade algorithm (see figure 5); (ii) Blocks and patches for extraction and processing of low-level features and statistical parameters: e.g., contraction index, directness index, stability index, quantity of motion, fluentness, pause and motion durations; (iii)Blocks and patches for posture recognition using both Hu moments (Hu, 1962) and simple pattern matching techniques. Posture recognition enables to associate postures to pause phases. Body postures and postural attitudes can have an important role in conveying expressive content and intentions to the audience and can be seen as a particular kind of expressive gestures. Argyle (Argyle, 1980) stresses the importance of postural attitudes in non-verbal communication: postures are used to express interpersonal attitudes, emotions, and personality traits, and they may be connected to speech and support it; (iv) Blocks and patches for analysis of movement in Laban's General Space: e.g., position in the General Space, expressive potential fields, occupation rates (Camurri, Mazzarino, Trocca, Volpe, 2001); (v) Patches for segmentation of movement in pause and motion phases. 5. CONCLUSIONS AND FUTURE WORKS This paper illustrates a methodology to face the problem of KANSEI analysis in human movement. This methodology is the application to human movement of a wider conceptual framework dealing with expressive gestures especially in their artistic manifestations, i.e., in interactive systems involving music, dance, visual media. In this perspective, this work should be considered as part of a broader research context in which KANSEI is seen as the deep level in which integration and mapping strategies are enabled. Our aim is to contribute to better support cross-language interactions, and enhance human computer communication. As a consequence, if from one hand future research will aim to better understand KANSEI by widening the set of cues to measure and developing cross-modal cues (e.g., based on comparisons of Schaeffer morphology with Laban's Effort theories), building suitable representations for expressive gestures, developing algorithms correlating concepts (Layer 4) to measured values of lower level cues. In particular, cross-modal integration (i.e., to relate these findings with similar analysis in other domains such as music and visual media), and to mapping strategies (i.e., the possibility for an automatic system to synthesize suitable expressive outputs depending on the analyzed KANSEI) is still needed. See for example figure 6 where simple but perceptually relevant movement cues are mapped on corresponding visual cues enabling a significant speed-up in the study of microdances with dancers and choreographers. The EyesWeb Motion Analysis library have been used in several public events, including interactive concerts (e.g. Roberto Doati work presented at La Biennale season, Sept 2001), and in museum installations (e.g. permanent interactive exhibits at Citta della Scienza science center, Napoli). ACNOWLEDGEMENTS We thank our collegues of the EyesWeb s/w development staff (Paolo Coletta, Massimiliano Peri, Andrea Ricci), the student Barbara Mazzarino, Eidomedia and NumenSoft. REFERENCES Argyle M, Bodily Communication, Methuen&Co Ltd, London, 1980. Camurri A., De Poli G., Leman M., and Volpe G., A multi-layered conceptual framework for expressive gesture applications, in Proc. Workshop on Current Research Directions in Computer Music, Barcelona, November 2001. Camurri A., Coletta P., Peri M., Ricchetti M., Ricci A., Trocca R., Volpe G., A real-time platform for interactive dance and music systems, in Proc. Intl. Conf. ICMC2000, Berlin, 2000. Camurri, A., Mazzarino, B., Trocca, R., Volpe, G. (2001). Real-Time Analysis of Expressive Cues in Human Movement. Proc. CAST01, GMD, St.Augustin-Bonn, Sept 2001. Cowie R., Douglas-Cowie E., Tsapatsoulis N., Votsis G., Kollias S., Fellenz W. and Taylor J., Emotion Recognition in Human-Computer Interaction, IEEE Signal Processing Magazine, no. 1, January 2001 Hashimoto S., KANSEI as the Third Target of Information Processing and Related Topics in Japan, in Camurri A. (Ed.) "Proceedings of the International Workshop on KANSEI: The technology of emotion", AIMI (Italia Computer Music Association) and DISTUniversity of Genova, Genova, 1997, pp 101-104 Hu, M.K. Visual pattern recognition by moment invariants. IRE Trans. Information Theory, vol. IT-8, pp. 179-187, Feb. 1962. Johansson G (1973) Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14, 201-211. Johnson-Laird P. M., Mental models and probabilistic thinking. Cognition, 50, 1994. Laban R., Lawrence F.C., Effort, Macdonald&Evans Ltd. London, 1947. Laban R., Modern Educational Dance, Macdonald & Evans Ltd. London, 1963. LeDoux J.E., The Emotional Brain, New York: Simon & Schuster, 1996. Schaeffer, P., Traits des Objets Musicaux. Second Edition. Paris. Editions du Seuil, 1977. Suzuki K., Hashimoto S., Modeling of Emotional Sound Space Using Neural Networks, in Camurri, A. (Ed.) Proc. Intl Workshop on KANSEI: The technology of emotion, AIMI and DIST-University of Genova, Genova, 1997, pp 116-121. Wallbott H.G., The measurement of Human Expressions, in Walbunga von Rallfer-Engel, "Aspects of communications", pp. 203-228, 1980. Open Computer Vision Library, Reference Manual. Available at: http://www.intel.com/research/mrl/research/opencv/ and http://sourceforge.net/projects/opencvlibrary/ 427

Page  00000428 Figure 5a: An EyesWeb patch showing the color blob tracker (dev by M.Peri and A.Ricci). Figure 5b: LK Trackers. i~t~r~rga i x-r J~lco........................~icto~ Figure 6: EyesWeb also supports mapping strategies, e.g., of movement into sound and visual outputs. In the example shown in the figure, the "quantity of movement" and "fluentness" cues are simply mapped in real-time on "intensity of light" and "color", respectively. We use similar representations - for example - for an intuitive and fast evaluation by dancers of movement quality. 428