Automatic Geo-registration and Inter-Sensor Calibration in Large Sensor Networks

Modern automated video analysis systems consist of large networks of heterogeneous sensors with different output characteristics, for example, static surveillance cameras, Pan-Tilt-Zoom (PTZ) cameras, infrared cameras, radars, and hyperspectral sensors. These systems must not only perform content extraction on individual sensors but also integrate and fuse the information from different sensors to effectively provide site-wide situational awareness. A critical step to analyzing and fusing data for site-wide scene understanding is to map observations from multiple sensors to a common coordinate system. In this chapter, we present a data-driven approach to automatic and semi-automatic estimation of inter-sensor mapping, relative topology of overlapping cameras, and geo-registration in large visual sensor networks. Data-driven approaches use sensor observations (for example, pairs of simultaneous target detections) over time to infer sensor geometry and network topology. Since these methods use target observation as their primary feature, they not only enable sensor registration in feature-less terrains but are also agnostic to the sensor’s output characteristics. In addition, they may also automatically adapt to the changes in sensor geometry. The data-driven approach presented in this chapter exploits domain and model-specific properties to develop efficient sampling-based mechanism for robust model estimation in the presence of outliers.