Distributed Smart Cameras and Distributed Computer Vision - Signal Processing Systems - page 463

calibration; these points can be selected using a variety of methods, such as SIFT.
Two cameras have overlapping fields-of-view if they can both see a certain number
of the X s.
The reconstruction is ambiguous and we need to estimate a matrix H that turns
the reconstruction into a metric form (preserving angles, etc.). A bundle adjustment
step uses nonlinear minimization to improve the calibration. Synchronization, also
known as temporal calibration, is necessary to provide an initial time base for
video analysis. Cameras may also need to be resynchronized during operation.
Even professional cameras may not provide enough temporal stability for long-
running experiments and low-cost cameras often display wide variations in frame
rate. Lamport and Melliar-Smith [ 16 ] developed an algorithm for synchronizing
clocks in distributed systems; such approaches are independent of the distributed
camera application. Velipasalar and Wolf [ 29 ] developed a video-based algorithm
for synchronization. The algorithm works by matching target positions between
cameras, so it assumes that the cameras have previously been spatially calibrated.
A set of frames is selected for comparison by each camera. For each frame, an
anchor point is selected. The targets are assumed to be on a plane, so the anchor
point is determined by dropping a line from the top of the bounding box to
the ground plane. Given these reference points, a search algorithm compares the
positions of the targets to minimize the distance D 1 , 2
i

between frame sequences 1

, j

and 2 at offsets i and j .
Pollefeys et al. [ 21 ] developed an algorithm for combined spatial and temporal
calibration of camera networks. They use a moving target, such as a person, to
provide data for calibration in the form of boundary points on the target's silhouette.
They use RANSAC to match points from frames from two videos. The RANSAC
search evaluates both the epipolar geometry and the temporal offset between the
frames. They first generate a consistent set of fundamental matrics for three cameras.
They then add cameras one at a time until either all cameras have been calibrated or
the errors associated with a camera's calibration are too large.
Porikli and Divakaran [ 22 ] calibrate color models of cameras by recording
images of identical objects from each camera and then computing a correlation
matrix of the histograms generated for the object from each camera.

6 Tracking

6.1 Tracking with Overlapping Fields-of-View

Velipasalar et al. [ 28 ] developed a peer-to-peer tracking system that fuses data
at a higher level. At system calibration time, the cameras determine all other
cameras that share overlapping fields-of-view. This information is used to build
groups for sharing information about targets. Each camera runs its own tracker.
As illustrated in Fig. 4 , a camera maintains for each target in its field-of-view the
position and appearance model of the target. The cameras then trade information