Advancements in sensing, processing, communications, and power technologies have enabled vast numbers of inexpensive wireless cameras that can observe a wide area. This dissertation addresses the challenges of designing a wide-area camera network and finding the spatial and temporal relationships between the cameras. In particular, we examine the challenges of calibrating non-overlapping cameras, aligning unsynchronized cameras with variable frame rates, and synchronizing cameras with variable frame rates. This research is largely motivated by the Scalable, Large Optical Sensor Network (SCALLOPSNet), a wide-area network deployed across the University of California, Santa Barbara. This dissertation describes the features of this network and the data that is used to validate our solutions. The first challenge is the 3D calibration of cameras with non-overlapping views.

We propose to calibrate each camera in a global geographic coordinate system that is introduced by a mobile device, such as a smartphone. We formulate this geo-calibration as a maximum-likelihood estimation of the position and orientation of each camera given multimodal sensor measurements from a mobile device. The mobile devices record multiple images that overlap with the camera to be calibrated with its GPS location and 3D orientation. We show that this estimation can be solved using a consensus algorithm. The second challenge is the alignment of overlapping cameras that have unsynchronized start times and variable frame rates. We use the intersections of object trajectories as a reliable feature point that can be matched across views. The potential matching intersections are derived from matching trajectories.

Trajectories are matched by, first, representing the spatial and temporal relationships between trajectories in a view in a Spatio-Temporal Context Graph and, then, matching these graphs across views. The final challenge is the synchronization of overlapping cameras with variable frame rates. Existing methods address cameras with fixed frame rates and known homographies. We formulate the problem mathematically to include variable frame rates, clock drift, and clock skew. The proposed solution is based on matching intersections of object trajectories. Addressing these challenges advances the state of the art in deploying wide-area camera networks. In particular, it enables the deployment of wireless cameras with non-overlapping views or unsynchronized frames.