Introduction
Surveillance and monitoring based on sensor networks have mainly consisted, up to now, in reading scalar magnitudes and sending them over to a relatively remote base station [1], [2], [3]. Other sensory modalities implying data structures of a higher complexity, like vision, have not been explored so far. Despite being the one of the most useful human senses [4], we have not been able to provide vision capabilities to sensor networks. At least, not without compromising any of the properties that make sensor networks so interesting, namely: autonomous operation, low deployment and maintenance costs, system scalability and adaptation to geo- and orographic characteristics of the area under surveillance [5].
Nevertheless, the interest in smart camera networks is increasing [6]. Embedded vision systems will be of major impact in consumer electronics in the near future [7] and powerful industrial alliances have been created in order to develop autonomous systems that are able to capture and interpret the visual stimulus [8]. There are applications related with surveillance and security, but still restricted to environments in which access to the power mains and node communication are granted a priori [9]. In places where infrastructure is poor or nonexistent, the major difficulty is to provide energy autonomy to the nodes. In other words, transfer them the capability of capturing, processing and transmitting multidimensional signal —like images or video— without compromising the useful lifetime of the network. If the node were able to interpret the scene and find out if whatever happens within the visual field is of interest, it would the transmitting irrelevant data and its associated cost. The problem would be to analyze the scene with a power budget smaller than the one required for continuously broadcasting the image sequence.
In connection with the local processing capabilities of a smart camera network, new functionalities emerge. For instance, by using local processing and inter-node communication, distributed collaborative and co-operative vision algorithms can be developed. Each node will share high-level information about the elements found within its visual field. Tracking without occlusion [10] and the interpretation of complex events by using information obtained from different points of view [11] will therefore be possible.