Graphics Reference
In-Depth Information
number of vistermswill have very long image chains. They also performed distributed
search: queries were distributed to the nodes, which computed a local ranking
of search results; these results were sent back to the proxy, which normalizes the
local search results by adjusting based on the number of images in which each vis-
term occurs and the total number of images in the database.
11.6 Cloud-Aware Systems
Although peer-to-peer is the clear choice for many algorithms, cloud-based computer
vision also makes sense in other applications. Both online analysis and search are
potentially amenable to the cloud. This is particularly true when we see peer-to-
peer and cloud as part of a continuum. Our challenge is to partition an algorithm
between camera nodes, the network, and the cloud. A variety of commercial services
provide cloud-based online analysis. Several retail customer analysis companies, for
example, use network cameras to ship video to the cloud for analysis. Widen [
30
]
surveys U. S. law on privacy and surveillance.
Law enforcement has received a great deal of attention as an application of both
online analysis and search. An episode of Nova [
29
] surveyed the various tech-
nologies, including video search, used to track down the Boston Marathon bombing
suspects. Much of the video used in the initial analysis was gathered by detectives
going door-to-door to local businesses. They also described NewYork City's Domain
Awareness System, which analyzes in real-time video from 4,000 video cameras as
well as environmental sensors and license plate readers. The system reads license
plates in every lane of the bridge and tunnel connections into lower Manhattan and
compares them to terror watch lists. It analyzes video for suspicious behavior in real
time; it can also perform query-based search.
We are also starting to see computer vision implemented as cloud applications.
CloudCV [
2
] provides a library of computer vision algorithms as a cloud service. It
provides Matlab and Python interfaces to its services. At the time of this writing, it
implements image stitching and object detection.
11.7 Conclusion
Advances in camera platforms are driving important new applications of distributed
smart cameras. Moore's Law provides not only high-performance computing nodes
but also advanced image sensors. FPGAs also provide rich platforms for the develop-
ment of custom vision platforms. Middleware can allow vision algorithm designers
to concentrate on vision without worrying about managing communication, mem-
ory, and processors. Peer-to-peer algorithms provide a powerful mechanism for the
design of ubiquitous smart camera networks. Cloud-aware vision systems comple-
ment peer-to-peer algorithms.
