A Distributed Approach to Global Semantic Learning over a Large Sensor Network

Sensor networks play an increasing role in several disciplines like security surveillance or environmental monitoring. A challenge of such systems is integration of observations from individual nodes into a common semantic representation of their environment. In this paper, an approach for establishing such a global view is presented, solely by correlating local information learned at the level of each sensing unit. Each node has two modes of processing environment information: it learns the “usual” events and “objects“ in its environment by means of probabilistic methods and it uses a set of predefined rules to detect standard situations that are uncommon or possibly dangerous. In the ongoing implementation of a security surveillance system, models of local activity at each node are inferred by processing video and audio data from its sensors, which are, for our test application, a camera and a microphone array. A local processing stack at each node serves to reduce spurious data from the sensors, to learn typical behavior and to detect unusual behavior by means of statistical techniques as well as predefined alarm situations by means of rule-based approaches. Communication with the neighbor nodes by means of a wireless network is used to build a shared understanding of the environment. The correlations of events and modalities among nodes are learned in order to establish neighborhood correspondences. The shared knowledge between neighbor nodes is used to enrich the local description of the learned normality. This can be used to establish paths over sensing zone boundaries of individual nodes, obtaining inter-node trajectories of observed objects. In addition, this approach avoids the necessity of precise calibration of the individual nodes.