Arena newsletter - October 2013
Transcription
Arena newsletter - October 2013
Architecture for the Recognition of Threats to Mobile Assets Using Networks of Multiple Affordable Sensors Arena N ewsletter Issue No 2, October 2013. A BRIEF PRESENTATION The EU FP7 project ARENA addresses the design of a flexible surveillance system for detection and recognition of threats towards deployment on mobile critical assets such as trucks, trains, vessels and oil rigs. The objective of ARENA is to develop methods for automatic detection and recognition of threats, based on multisensory data analysis. MID-TER M REV IEW On the 21st of March, in Stockholm Sweden, the ARENA project was under review as part of the EU evaluation process. The review meeting was successful with many interesting discussions and questions. The project reviewers concluded that there is good progress within ARENA and that the project has achieved most of its objectives and technical goals for the period with relatively minor deviations. The objectives are still relevant and the mid-term review concluded that the objectives are also achievable. ■ DISSEMINATION ACTIV ITIES Truck protected by the ARENA system The 10th IEEE International Conference on Advanced Video and Signal-Based Surveillance (AVSS) was held in Krakow, Poland, on the 27-30th of August 2013. In this conference, the ARENA project represented itself with a contribution entitled ‘Activity recognition and localization on a truck parking lot‘. The contribution, jointly prepared among the partners, was presented at one of the conference’s poster sessions. For more information on the conference, visit: www.avss2013.org The full citation of ARENA’s contribution is: Andersson, M., Patino, L., Burghouts, G., Flizikowski, A., Evans, M., Gustafsson, D., Petersson, H., Schutte, K., Ferryman, J., “Activity recognition and localization on a truck parking lot”, The 10th IEEE International Conference on Advanced Video and Signal-based Surveillance (AVSS 2013), Krakow, 27-30 August, pp. 263-269, 2013. ■ Vessel protected by the ARENA system. The final project results will be demonstrated in a live demo and pre-recorded demos in Paris in April 2014. The demonstrations will show the principles of the ARENA system in a truck scenario, and optionally in a maritime scenario. ARENA has a stakeholder group with representatives from both the land and maritime cases. The stakeholder group has played an important role in the development of user requirements, specifications and scenario definition. ■ COLLABOR ATIVE ACTIV ITIES Data collection campaign (no. 2) A data collection campaign was arranged at the University of Reading, UK, early September this year. The purpose of this additional campaign was to collect more video surveillance data to further evaluate all algorithms developed within ARENA. Integration Workshop A two day workshop was held in Paris on the 3-4th of October. The purpose of this workshop was to gather all Architecture for the Recognition of Threats to Mobile Assets Using Networks of Multiple Affordable Sensors A few photos from the data collection campaign undertaken at University of Reading, UK, Sepetember 2013. partners for hands-on work with the integration platform. The meeting was fruitful and the project made progress towards the final demonstration. ■ now begin to settle. The system will incorporate all levels from low-level sensor data processing to high-level decision support and HMI-interfaces. Recent and state-of-art scientific results are used at all these levels. TECHNICAL OVERV IEW Object Detection and Tracking ARENA integration platform A backbone of the ARNEA-system and the test bed that is going to be presented at the end of the project is the integration platform(IP). It is a challenge to set up a complete sensor system with effective and modular approaches to communicate information between all system components. The IP will provide an effective mean to provide all algorithms with sensor data and to make results easily accessible for operators etc. Currently, intensive and collaborative work is ongoing to test and make minor adjustments to the implemented IP. The finalized IP will make an important outcome of the ARENA project that can be used as a foundation when designing other, ARENA-like, systems. Algorithm Development Not only the integration platform, but also the technologies and methodologies used within the ARENA system Object detection and tracking is used to detect and track interesting objects, e.g. pedestrians, within a scene. Approach for detection The background areas within each camera view is adaptively described using a Gaussian Mixture Model. All model parameters are updated online (Zikovic, ICPR’04). By using the background description, all foreground elements can be extracted. Pedestrians can then be detected by using a classifier to classify all foreground elements. The ‘Fastest Pedestrian Detector in the West’ is used for this task (Dollar, BMCV’10). Approach for tracking Standard methods for organizing the sequence of detections (above) into consecutive tracks are applied: • Linear Kalman filters • Constant velocity motion models • Multi-hypothesis tracker for data associations. Architecture for the Recognition of Threats to Mobile Assets Using Networks Trajectory speed changing points of Multiple Affordable Sensors Initial set of zones Learned zones Zone learning methodology: (i) Multi-resolution analysis of all mobiles speed profile to extract speed changing points. (ii) Speed changing points are the input to a fast clustering algorithm (Leader, Duda et al. 1995). The clustering results in an initial set of zones {Zn}. (iii) The partition of clusters is corrected by merging similar zones, Zn, employing soft-computing relationships. Early results from algorithms performing object detection and tracking Action Recognition The ARENA-system analyses imagery from detected pedestrians and classifies their actions into simple categories such as {walk, run, turn, check, fight, enter, loiter}. Approaches for action recognition The action recognition method is effectively one detector for each action (Burghouts & Schutte, ICPR’12). Each action detector quantifies STIP features by a soft-assignment random forest (Burghouts, IJPRAI’13). Locations of the motion features in a 3D volume are captured by a Gaussian layout model (Burghouts & Schutte, PRL’13). The bag-of-words histograms are classified by an actionspecific SVM (Burghouts, Schutte, Bouma & den Hollander, Machine Vision and Applications’13). Group and Fight Detection Groups and fights will be detected by analysing how densely pedestrians are located within a scene. walk 80 2 4 6 4 2 2 run 21 64 0 7 0 0 7 loiter 4 0 78 12 4 2 0 turn 6 4 16 62 4 4 4 enter 8 0 0 10 82 0 0 Zone Based Events check 0 0 6 19 6 63 6 fight 13 3 0 10 7 0 67 run loiter turn enter check fight Approaches The group detection is based on K-means clustering (Hastie et al, The elements of statistical learning, 2009), the silhouette value (Rousseeuw, JCAM’87) and the group density measure. Group detection consists of three steps (Andersson, Gustafsson, St-Laurent, Prévost, JSTSP’13): (i) Tracking points are clustered for segmenting people into clusters. (ii) The K-value yielding the highest silhouette value gets to represent the current number of clusters. (iii) For that K we calculate the group density measures to see if there are any dense clusters. walk true activities Classification performance of human activities on a parking lot (in percentages). Overall performance 70.8%. Top-view of a parking-lot. User defined areas are indicated with arrows (T=truck, SZ=smoking area, SA=service area, PW=?). Learned zones are in red. Many of the real-world events of interest for the ARENA system can be described by movements within and between different zones within a scene. predicted activities Automatic zone learning Activity zones of the scene are those areas where people interact or perform behavioural changes: stop, walk to meet someone, speed up walking, or stand waiting. Zone based events Pedestrians patterns of transitions between activity areas can be learned automatically and used to define situational events. A typical delivered event can be: “From just south of zone Truck to just north of service area”. Architecture for the Recognition of Threats to Mobile Assets Using Networks True positive rate (TPR) for recognition of zone-based events using data from the Paris data collection campaign . Events Instances TPR From smoking area to car 1 100% From car to smoking area 1 100% From car to truck 3 0% From service area to truck 11 82% From truck to car 2 0% From truck to service area 11 91% of Multiple Affordable Sensors ON THE AGENDA Intensive work and efforts will be made during the autumn of 2013 to make progress towards the final implementations and to test the integration platform. New hands-on workshops are being planned for and the planning of the final demonstration is ongoing. ■ Ontology Support An ontology is included in the design of the ARENAsystem. The purpose of the ARENA Ontology is (i) to enable automatic configuration of threat recognition algorithms given e.g. current conditions and camera position, etc. (ii) It is also to support situation assessment of the parking area – e.g. by indicating threats and properties of elements in a parking zone. FURTHER INFOR MATION Illustration on information that could be provided by the ARENA ontology: parking’s elements (static and dynamic), relations and dependencies between parking’s elements, threats, properties. Further work (on algorithm) Further R&D work include to improve the tracking procedure: With increased continuity in position estimates the event recognition performance can be further improved. Another area to investigate further is sensor fusion, and event fusion: By introducing fusion, events can be observed from different perspectives and the event recognition performance can be further improved. ■ The ARENA project is coordinated by FOI. For inquiries and requests for further information, please visit the project’s site on internet: www.arena-fp7.eu