Hermes, the French fashion brand, early in order to create a well-known senior horses in Paris, France, and after the introduction of the bags, clothing, scarves, perfume, enamel, jewelry and household goods, make the brand more comprehensive diversity. hermes is also the ancient Greeks called Mercury.
2 0 0 6 – 2 0 0 9 F i n a l P r o j e c t R e p o r t t · cf ES M ER H .4 1:1 m 50 m Hermes Human Expressive Representations of Motion and their Evaluation in Sequences http://www.hermes-project.eu hermes-1.indd 1 10/09/2009 19:37:14 hermes-1.indd 2 10/09/2009 19:37:14 The HERMES Project 3 “HERMES concentrates on how to extract descriptions of human behaviour from video sequences, in a restricted discourse domain.” We confront agent, body and The sensing and reasoning The system has been exposed to face motion evaluation, in sce- processes are tightly integrated video recordings from different narios ranging from wide-field- in a perception-action cycle: co- parts of Europe. Its explanatory of-view with multiple-agent operating pan-tilt-zoom sensors and arguing capabilities assess scenes to specific inferences of enhance cognition via respons- its strengths and weaknesses. emotional state. es to uncertain/ambiguous in- terpretations. Project Nº 027110 Duration 36 months (+6 months extension) Start Date 01/03/06 End Date 28/02/09 (extended to 31/08/09) European Commission Funding 2 100 000 , 00 € Hermes homepage http://www.hermes-project.eu Juan José Villanueva Project Manager H ead of H ermes Istanbul arcHaeologIcal museum , turkey hermes-1.indd 3 10/09/2009 19:37:21 4 Executive Summary Hermes Project Objectives HERMES is a consortium project that The main objective of HERMES is to de- concentrates on extracting descriptions velop a cognitive artificial system to al- of people behaviour from videos in re- low both recognition and description stricted discourse domains, such as pe- of a particular set of human behaviours destrians crossing inner-city roads, ap- arising from real-world events. Specifi- proaching or waiting at stops of buses cally, we model knowledge about the and even humans in indoor worlds like environment to suggest interpretations halls or a lobbies. from motion events, and to communi- These video recordings allow us to ex- cate with people using natural language plore a coherent evaluation of human texts and synthetic films. The events movements and facial expressions are detected in image data-streams across a wide variation of scale. obtained from arrays of multiple active cameras (including zoom, pan and tilt). A system has been developed that starts with basic knowledge about pedestrian The HERMES procedure combines: behaviour in chosen discourse domains, Detection and tracking of agents but eventually clusters evaluation re- while they are some distance away sults into semantically meaningful sub- from a location. sets of behaviours. If they are even closer and their face can be resolved, facial emo- tions are checked. “We base on the three stages of human behaviour: motion of people, their posture and gestures, and their facial characterization.” H ermes , sculpture of szeroka street toruń , poland hermes-1.indd 4 10/09/2009 19:37:27 5 “Natural language texts and synthetic animation are used to communicate with end-users” Approach Animation is accommodated within HERMES based on the following consid- HERMES fulfils two main objectives: on erations: the one hand, the generation of concep- Analysis-by-synthesis at the three tual descriptions based on acquired and stages of human behaviour, i.e. analysed motion patterns. On the other motion of people, their posture hand, the communication using visuali- analysis, and their face characteri- zation of synthetic motion patterns. zation. Natural language text generation is ac- Animated computer graphics as a commodated within HERMES based on visual language to quickly commu- these considerations: nicate essential aspects.. Semantic descriptions enable re- Animated computer graphics, searchers to check details of the again at three motion categories, conceptual knowledge base. for checking the underlying con- They also allow communication ceptual knowledge base. with end-users in a natural manner. Using both approaches, quantitative They support conceptual abstrac- measures and qualitative descriptions tion, thereby facilitating the com- have been developed to analyze the ro- munication of short messages or bustness and efficiency of the proposed essential details. cognitive system. sandalbInder H ermes untyIng HIs sandal louvre museum , france hermes-1.indd 5 10/09/2009 19:37:31 6 Contents Executive summary 4–5 Contents 6 HERMES Consortium 7 HERMES Workpackages 8–19 WP0: Project management 10 WP1: Definition of requirements 11 WP2: Distributed camera system 12 WP3: Evaluation of agent motion in video sequences 13 WP4: Evaluation of body motion in video sequences 14 WP5: Evaluation of face motion in video sequences 15 WP6: Software integration for generation of natural language texts 16 WP7: Software integration for generation of virtual environments 17 WP8: Validation and demonstration 18 WP9: Exploitation and dissemination 19 Showcasing Activity 20–21 Expertise of the HERMES partners 22–23 Finances 24 c aduceus (Wand of H ermes) hermes-1.indd 6 10/09/2009 19:37:35 HERMES Consortium 7 Computer Vision Center CVC Spain Universitat Autònoma de Barcelona Institut für Algorithmen und Kognitive Systeme IAKS Germany Universität Karlsruhe Computer Vision and Media Technology Laboratory CVMT Denmark Aalborg University Computer Vision Laboratory BIWI Switzerland ETH Zürich Active Vision Laboratory AVL United Kingdom University of Oxford Answare Technologies Spain hermes-1.indd 7 10/09/2009 19:37:44 8 WP0 Project Management Project Manager Juan José Villanueva Steering Management and Committee Exploitation Office Technical Coordinator Objectives Andrew D. Bagdanov The objective of WP0 is the project or- ganization and the management of the project activities. The main objectives Management are to initiate the project phases, to Coordinator complete all the necessary administra- tive tasks required by the Commission, Tonny Velin and to provide short progress reports and the agreed deliverable results to the EC on due time. Scientific Coordinator Jordi Gonzàlez Work Package leaders WP0 WP1 WP2 WP3 WP4 WP5 WP6 WP7 WP8 WP9 WP0–1 leader Juan José Villanueva email@example.com WP leader: WP leader institution: hermes-1.indd 8 10/09/2009 19:37:44 9 Definition of requirements WP1 Objectives WP1 synthesizes the requirements that define the technology to be developed in the context of HERMES. Due to the risk inherent to the complexity of the developed system, the work in WP1 es- tablishes a realistic set of requirements, assumptions and restrictions to fulfil the Agent Body Face project objectives and also to allow po- Motion Motion Motion tential applications (of interest to part- ners) while taking into account the lim- ited resources allocated. Activities & achievements The defined requirements were elabo- rated, completed, synthesized, and harmonized to assure a complete and Human expressive representations of motion... consistent output for use in the devel- opment, design, and implementation phases. An ontology was also defined to determine what the system is even- tually going to say; which videos were Validation to be recorded; what had to be select- ed, tracked, and converted into natural language output; and which questions were to be answered by the system. From Images From Texts To Texts To Images Since the output of this WP was input for all other WPs, it detailed and con- textualized the global requirements for each workpackage scope. The discourse domains were specified and also the locations for recording human agents. The considered scenarios provide a common test input for all partners, but deliberately allowing them for leeway to record additional videos for testing. ...and their evaluation in sequences hermes-1.indd 9 10/09/2009 19:37:44 10 WP2 Distributed Camera System Objectives Activities & achievements We made progress on information theoretic means to control a set of WP2 provides a platform to acquire video In addressing these two goals, the con- cameras. The key idea is that lim- data for use in WP3, WP4, and WP5, and sortium has contributed on three ways: ited sensing resources should be investigates the link between high-level directed to where they will yield the We developed an architecture for highest “information” gain. understanding and sensing action. Its multi-camera systems comprising goals are: to provide an architecture from which to serve video data for WP3, WP4, asynchronously communicating, Future directions heterogeneous cameras. An SQL and WP5; and to couple high-level scene We have just scratched the surface of pos- database is used for data archiving interpretation with sensing actions. sibilities in the feedback from high-level and inter-camera communication. inference to sensing process. An imme- We interfaced the multi-camera diate step will be to link our information system to WP6’s inference engine theoretic control into a live system as an and collaboratively created Situa- internediate layer between inference and tion Graph Trees to describe inter- action. We also expect our ideas to ex- esting activities in scenes. tend naturally to other cognitive systems. HERMES multi-camera architecture Collaborative distributed cameras (static and active) WP2–3 leader WP Reid Ian D.leader: firstname.lastname@example.org WP leader institution: hermes-1.indd 10 10/09/2009 19:37:45 11 Evaluation of Agent Motion in Video Sequences WP3 Tracking by segmentation Objectives reasoning about the scene. This infer- ence has, in turn, been used for natural WP3 aims to track agents and note their tra- language generation (WP6), generation jectories and other coarse scale features, in of virtual agents (WP7), and camera selec- order to generate conceptual descriptions tion/control (WP2, WP8). about the agents and their relationships among them and with other scene objects. Future directions Activities & achievements Incorporating high-level knowledge with low-level tracking data is challenging. At the heart of most surveillance systems Current tracking systems typically operate is the requirement to track targets. We in states of perpetual surprise; algorithms developed new algorithms for figure- try hard not to lose the target, and little ground segmentation-based tracking, intelligence is used in reacquistion. By con- and integrated a real-time implementa- trast, if an activity pattern emerges, a sys- tion into the final demonstrator system. tem –like a human– can afford periods of We have also been at the forefront of deliberate inattention, having a sufficiently tracking-by-detection methods, publish- rich model to be confident of reacquisition. ing at IEEE CVPR, ICCV, and BMVC, and Enumerating all possible situations in a winning a best-paper prize at CVPR 2007. SGT rapidly becomes infeasible for even WP3 is about more than tracking: close small numbers of agents or vaguely com- collaboration allowed us to integrate an plex scenes. A key challenge will be to inference engine with the agent-based capture scene knowledge by learning, in- Multiple target tracking using person tracking data, to conduct simple real-time stead of hand-crafting the SGTs. detectors as input hermes-1.indd 11 10/09/2009 19:37:45 12 WP4 Evaluation of Body Motion in Video Sequences Activities & achievements WP4 activities concentrate on: collecting relevant test data sets, pose estimation, and action recognition. Pose estimation is relevant on its own (e.g., attentional focus via head pose), and also a basis for action recognition. A multitude of actions such as waving, Head-pose estimation boxing, or running are recognized by Objectives algorithms that have different foci re- garding their goal: view-invariance, real- This WP is the core of action evaluation: time, recognition rate, etc., leading to already knowing the rough position of distinct approaches based on features, the body, derived from WP2 and WP3, it silhouettes, or depth. The algorithms supports both the recognition and the closely integrate with WP3’s segmenta- animation of body postures. The pose tion and tracking, yielding comprehen- and motion of the body is modelled, es- sive systems. timated, and associated to verbs such as walking, standing, or bending. Future directions Recognized action: ‘Running’ Future methods on segmentation and tracking should be more robust against partial occlusion, background clutter, and problematic lightning conditions. It is also a challenge to enhance robust- ness by combining multiple cameras and using depth data. Lastly, context in action recognition has shown to be a promising avenue to follow. WP4 leader Thomas B. Moeslund WP leader: email@example.com WP leader institution: hermes-1.indd 12 10/09/2009 19:37:50 13 Evaluation of Face Motion in Video Sequences WP5 Objectives Several tools and methods for 2d and 3d face tracking and face alignment have The focus of WP5 is on the develop- been developed that allow for auto- matic processing of facial motion in real- ment of algorithms for facial expression time. The systems are designed to proc- processing and the analysis of facial mo- ess a wide range of input streams from tion in continuous image streams. The low-resolution 2d image data to 2.5d main issues that we addressed are au- data acquired by active or passive stereo tomatic face processing, real-time per- systems. A head pose tracker has been formance, and several levels of detail to developed in collaboration with WP4. meet the various requirements of end- Several levels for the generation of facial users. Other aspects are the modeling motion descriptions have been consid- and generation of emotion descriptions ered, ranging from low-level speaking which serve as input for the animation and non-speaking classification over of virtual faces. speech recognition to high quality facial emotion estimation. Activities & achievements In cooperation with WP7, a virtual face model has been designed based on the MPEG-4 standard to visualize the gen- A system for emotion modeling has erated emotion descriptions. Further- been specified for distributed camera more, a system has been developed that systems (WP2) with pan-tilt-zoom sen- accurately aligns a 3d template face to sors controlled by the tracked agent mo- 2.5 data. By pre-recording the basic fa- tions (WP3). For evaluation, a database cial expressions, the system tracks and containing image sequences of human transfers the expressions of a human to emotions has been built and made avail- a facial avatar in real-time. able to the scientific community. Face alignment for the analysis of 2d face images Recognition and transfer of an expression to a facial avatar WP5 leader Luc Van Gool WP leader: firstname.lastname@example.org WP leader institution: hermes-1.indd 13 10/09/2009 19:37:50 14 Software Integration WP6 for Generation of Natural Language Texts 3@630 A pedestrian labelled as ‘actor 3’ appears in the field of view. 3@632 He moves on the southeastern sidewalk. 3@663 He walks within it. 3@766 He walks on the zebra. 3@900 Actor 3 stands near by another pedestrian. 3@928 Actor 3 moves on the zebra. 3@1085 Actor 3 stands near by another pedestrian. 3@1147 Actor 3 moves on the zebra. 3@1250 He walks on the street. 3@1254 Actor 3 chases another pedestrian. 3@1289 Actor 3 runs on the street. A pedestrian chases another pedestrian. Objectives demonstrator system generates natural language text on activities of a particu- Imagine a user communicating with a set lar agent (human or road vehicle) from of distributed PTZ-cameras, as if they were a schematic conceptual representation humans reporting what they see. This re- expressed as a set of Fuzzy Metric-Tem- quires converting a video into a textual poral Horn Logic formulas, based on the description of temporal events. The user agent’s trajectory data. should be able to request summaries of recent developents in chosen languages; Future directions to obtain responses to his questions for details; and to send commands, e.g., to The capability to detect, represent, and zoom in onto a particular body. describe more involved agent behavior is desirable, including stylistically ad- Activities & achievements equate formulations for a larger range The HERMES-consortium has designed, of behaviors. Also, a combination of the implemented, and tested various pro- algorithmic rigor associated with a for- totypical systems to cover the differ- mal logic approach and the flexibility ent requirements set out. Currently, a expected from a ML-oriented approach slimmed-down, fixed- and PTZ-camera poses a challenge for future research. Pedestrian crossing a road via a traffic island. 0@210 A pedestrian labelled as ‘id 0’ comes in from the north sidewalk. 0@220 He stops. 0@222 He walks on the north sidewalk. 0@231 He stops. 0@234 He stops to due a car labelled as ‘id 1’. 0@239 He walks on the north sidewalk. 0@266 He walks on the crosswalk. 0@371 Now he has reached the center island. 0@401 He walks on the crosswalk. 0@503 He walks on the south sidewalk. WP6 leader Hans–Hellmut Nagel WP leader: email@example.com WP leader institution: hermes-1.indd 14 10/09/2009 19:37:51 15 Software Integration for Generation of Virtual Environments WP7 Objectives Producing virtual human motion im- plied animating the characters with The goal of WP7 is to visualize the con- the motion descriptions used thorough ceptual descriptions generated for be- WP3-5. The integrated motion was ex- amined for consistency using param- haviour (WP3), action (WP4) and emo- eters of emotion and style, to find stable tion (WP5), joining them into a common links between both kinds of animations. animation framework. WP7 generates Particularly, we investigated whether a synthetic animations reproducing se- given emotion may influence the anima- quences of predefined semantic primi- tion of particular actions or behaviours. tives, provided beforehand or derived Additionally, generating virtual anima- from end-user’s textual descriptions. tions based on conceptual descriptions can be seen as inverting WP6’s natural Activities & achievements language generation: here, the seman- tic primitives related to behaviours, ac- Producing virtual components includes: tions, and emotions are visualized us- ing computer graphics. Since a virtual Conception, creation and visualiza- environment contains all motion and tion of a virtual scene; position information on all agents and Its texturing and illumination; objects, interactions can be revisited. This system was evaluated for accuracy Virtual actors generated from users’ 3D-modeling of virtual objects and of visualization and degree of interac- descriptions and reacting to real scenes. human characters; and tion. As a result, behaviour descriptions Their inclusion into the virtual scene. are visualized by means of virtual actors. WP7 leader WP Gonzàlez Jordi leader: firstname.lastname@example.org WP leader institution: hermes-1.indd 15 10/09/2009 19:37:52 16 WP8 Validation and Demonstration Objectives an outdoor scenario was designed to probe the effectiveness of the HERMES WP8 focuses the integration, demon- approach to image visual cognition in stration and validation of the other WPs’ multi-sensor networks. An integrated outputs. A major goal of HERMES is to hardware platform was designed, built establish how the analysis of agent, face and installed on the rooftop of the CVC and body motion, plus synthesis of these building in Barcelona. This hardware plat- outputs in the form of inferred, high-level form consists of two high-speed cameras interpretation should be integrated into (one fixed and one pan-tilt-zoom), and a cohesive platform for experimentation, three dedicated servers to host HERMES validation and demonstration. In WP8, the systems for analysis of agent motion, outputs of all WPs were pooled and inte- for active camera control, and for infer- grated to build integrated hardware and ring high-level descriptions of agent be- software systems to meet such objectives. haviour in the scene. To further support demonstration activity, an integrated and modular software platform was de- Activities & Achievements signed, allowing modules to be devel- oped off-site and later integrated into the In the final year of the project, the con- platform with minimal effort. sortium turned its attention in earnest to- Real-time tracking and analysis of agent wards realizing both real-time and offline motion results from WP3, active camera demonstration and validation systems. control from WP2 and inference of agent Building on the prototype indoor dem- behaviors from WP6 and WP7 have been onstrator developed in WP2 and WP3, incorporated into the demonstratiors. Real-time demonstrator interface controlling the fixed camera (top), active camera (bottom left), and split view of both cameras (bottom right). WP8 leader Andrew D. Bagdanov WP leader: email@example.com WP leader institution: hermes-1.indd 16 10/09/2009 19:37:59 17 Exploitation and Dissemination WP9 Objectives project. Presentations at national and international conferences and forums WP9 promotes the dissemination of the have been to be given (e.g. ICPR, CVPR, project results to the wide public, mak- ECCV, ICCV, and other international and ing them well known throughout in the national conferences and scientific jour- scientific community and the European nals, like IEEE Transactions on Pattern industrial companies. It also promotes Analysis and Machine Intelligence, Com- the industrial and market exploitation of puter Vision and Image Understanding, the project results. Pattern Recognition Letters, etc). The construction of the HERMES web- Activities & Achievements site, http://www.hermes-project.eu, has greatly contributed to disseminate Many interchanges of research people the project results. The website contains among the HERMES partners took place demonstrators of the project highlights during the whole project duration. The and other results supplied by the part- research partners have disseminated ners. It has also been used to archive re- the results of the project with publica- ports among the partners. tions and presentations in events and conferences. Technical papers and arti- The project partners will exploit the cles in scientific and professional jour- project results by using the basic princi- nals and specialized magazines have ples of HERMES in commercial projects been published demonstrating the work (e.g. smart surveillance for security that has been carried out during the projects) and other research activities. WP9 leader Tonny Velin WP leader: firstname.lastname@example.org WP leader institution: hermes-1.indd 17 10/09/2009 19:38:01 18 Showcasing activity “A hardware and software platform integrates the components of HERMES” HERMES showcase activities have been concentrated in the form of an integrat- ed hardware and software platform that consolidates the various components of the HERMES approach. In concert with the demonstration and integration activities of WP8, two inte- grated graphical user interfaces have been implemented to demonstrate and showcase the performance of HERMES modules: An interface for interacting with the real-time demonstrator was built that allows the user to configure, administer and monitor the com- ponents of the HERMES demon- strator platform. This interface also allows the user to visualize stream- ing video from both cameras in re- al-time, as well as tracker output in the form of 3D annotations on live video. Inference results from analy- sis and reasoning about agent mo- tion is also presented to the user in this interface, along with captured detail images from the PTZ camera. To complement the initial real-time demonstration system, a second offline demonstrator for anima- tion of virtual environments, based on the outputs of WP7, was con- structed that bases its virtual re- construction of agent behaviour on the outputs of the real-time demonstrator. The end result of these efforts is a flex- ible and modular system for showcasing some of the major developments from the HERMES project. H ead of H ermes from a HermaIc stele ancIent agora museum of atHens , g reece hermes-1.indd 18 10/09/2009 19:38:04 19 “Two demonstration systems, one real-time and one offline, showcase the major developments of HERMES” H ermes bearIng tHe good person by praxIteles arcHaeologIcal museum of ancIent o lympIa , g reece hermes-1.indd 19 10/09/2009 19:38:15 20 Expertise of the Partners CVC CVMT The CVC is a non-profit institution The Computer Vision and Media Tech- founded in 1995 by CIRIT (Consell Inter- nology Laboratory (CVMT) is part of the departamental de Recerca i Innovació technical faculty at Aalborg University, Tecnològica), CIDEM (Centre d’Innovació i Denmark. CVMT conducts research in Desenvolupament Empresarial) and UAB areas such as 1) computer and robot vi- (Universitat Autònoma de Barcelona). The sion, 2) computer graphics, 3) motion strategic objective of CVC is to do both analysis, 4) color vision, 5) autonomous Research and Development of quality systems and agents, 6) multimedia and on Computer Vision. From a scientific virtual reality (VR) systems and inter- point of view, the CVC wants to contrib- faces, and 7) operational vision systems ute to increase the knowledge in this for both industrial and medical applica- field. From a technological point of view, tions. CVMT has been involved in about the CVC aims to contribute to innova- 15 European projects and a number of tion and industrial competitiveness col- national projects. Currently CVMT is a laborating with companies to develop part of and coordinating one EU project technological projects. and three national project. IAKS BIWI The Fakultät für Informatik der Univer- At the Eidgenoesssiche Technische sität Karlsruhe (TH) is one of the lead- Hochschule (ETH) Zurich the participat- ing faculties for Computer Science in ing team is the Computer Vision Labora- Germany. The research group Kognitive tory (BIWI), which is part of the Dept. of Systeme (KOGS) lead by H.–H. Nagel is af- Electrical Engineering D-ITET. The work filiated with the ‘Institut für Algorithmen of the group focuses on scene under- und Kognitive Systeme’ der Universität standing, medical imaging, and remote Karlsruhe. This group addresses all prob- sensing. Here mainly the expertise in the lems encountered in the algorithmic area of scene understanding is relevant. transformation of image sequences into This includes work on tracking, gesture representations of the semantics of the analysis, and motion capture. The group depicted scenes and their temporal de- has developed multiple trackers, rang- velopment, covering data-driven treat- ing from generic, fast blob trackers up ment of video signals via model-based to trackers for articulated structures like approaches towards detection and hands or full bodies. The group has been tracking up to logic-based representa- involved in many European and Swiss tions and their conversion to natural projects. language text. hermes-1.indd 20 10/09/2009 19:38:20 21 AVL The Active Vision Laboratory is part of the Robotics Research Group in the University of Oxford’s Department of Engineering Science. Since its inception in 1991 the lab has established a long record of research in active, real-time and robotic vision. A substantial part of the group’s effort is invested in research- ing the theory of visual sensing and per- ception and its application to a wide variety of problems. The group has had substantial success in both visual track- ing and in 3D geometry. The group has been involved in a numerous previous European projects. Its faculty have con- sulted widely for major companies. ANSWARE Answare Technologies is a Spanish SME specialist in Information and Communi- cation Technology, providing consultan- cy and projects development services and offering a variety of high-tech solu- tions based on state-of-the-art technol- ogies, tailored to the specific needs of customers. Answare operates in nation- al and international market sector for Aerospace, Defence, Telecommunica- tions and Internet. Answare’s staff have participated in the complete life cycle for software systems projects from con- ception through development, delivery and into maintenance. H ermes IngenuI WItH kerykeIon (Herald’s staff), kItHara , petasus (round Hat), traveller ’s cloak and WInged temples pIo clementIno , vatIcan cIty hermes-1.indd 21 10/09/2009 19:38:23 22 Finances Hermes expenditure per year (of the 2.100.000,00€ grant) M€ 800 700 600 500 400 729.290 € 742.869 € 300 627.841 € 200 100 0 Period 1 Period 2 Period 3 Mar 2006 / Feb 2007 Mar 2007 / Feb 2008 Mar 2008 / Aug 2009 The HERMES project has been allo- cated a budget of 2.100.000 € by the European Commission to fulfil the objectives of its work programme. This grant was distributed over the 42 months period of the project (36 months + 6 months extension), from February 2006 to August 2009. The grant was divided among all the partners, according to their activities and commitment in the project. “The HERMES project has been allocated a budget of 2.100.000 € by the European Commission” H ermes WearIng tHe petasus , a voyager ’s cloak , tHe caduceus and a purse museo cHIaramontI , vatIcan cIty hermes-1.indd 22 10/09/2009 19:38:26 Authors: Written with the contribution of all HERMES Partners Designed by: Carles Fernández Tena hermes-1.indd 23 10/09/2009 19:38:26 http://www.hermes-project.eu September 2009 hermes-1.indd 24 10/09/2009 19:38:26
Pages to are hidden for
"Hermes brochure"Please download to view full document