SCOVIS will significantly improve the versatility and the performance of the current monitoring systems for security purposes and workflow control in critical infrastructures. The resulting technology will enable the easy installation of intelligent supervision systems, which has not been possible so far, due to the prohibitively high manual effort and the inability to model complex visual processes.
An automobile industry has been selected for the evaluation of the SCOVIS research tools under a real world environment. SCOVIS supports the automatic detection of a) behaviours, b) workflow violation and c) localization of salient moving or static objects in scenes, monitoring by multiple cameras (static or active).

The project investigates weakly supervised learning algorithms and self-adaptation strategies for analysis of visually observable workflows and behaviours. The goal of these algorithms is to use a relatively small number of labelled data, almost at the initial stage of the algorithm, while in the following, unlabelled data are exploited. Camera network coordination is also supported so that complex behaviours can be identified as combination of spatio-temporal object relations in multiple scenes. SCOVIS supports self-configuration (system is able to automatically calculate the camera spatial relations) and adaptation (the models are automatically enriched through time via online data acquisition and unsupervised learning strategies). User’s interaction is also foreseen for improving the behaviour detection through relevance feedback mechanisms. This way, the user evaluates the system performance and then the rules used by the system are automatically updated (without imposing any additional knowledge from the user about the system operation) so that in the following responses better decisions are accomplished. The proposed research will be performed with absolute respect to privacy and personal data of monitored individuals.

The expected outcome of the project comprises the following:

• A methodology for largely unsupervised learning of salient objects.
• A methodology and an open architecture for large-scale camera networks. It will define the methods for automatically setting up, operating and maintaining networks of active or static cameras with overlapping or disjoint views.
• A toolkit for weakly supervised learning and object detection. The toolkit will include tools for detection and learning of moving or static objects using a library of generic features and attention models.
• A toolkit for behaviour analysis. This will include tools for workflow learning and recognition for single or multiple agents. It will also include tools for workflow disambiguation.
• A toolkit for adaptation mechanisms. It will include tools for model adaptation (object or behaviour) using either automatic mechanisms or user feedback.
• A toolkit for camera network coordination. It will include tools for consistent monitoring using cameras with overlapping or disjoint field of view, and for controlling an active camera.
• A testbed which will integrate all above toolkits and will demonstrate the benefits of the proposed synergies.

Watch videos from JRS work on SCOVIS: person detection, active camera tracking, and camera calibration

YouTube Video Real-Time Detection of Unusual Regions in Image Streams

Project Homepage: http://www.scovis.eu