This document discusses using cloud computing for CCTV video surveillance. Some key points:
- Cloud infrastructure reduces the need for in-house IT resources and provides scalability, reliability and cost savings compared to on-premise systems.
- However, storing and transferring raw CCTV video data to the cloud is not viable due to the huge size of video files and bandwidth limitations.
- Instead, "smart cameras" can analyze video locally to extract metadata and detect meaningful events, only transferring thumbnail images or short clips to reduce transmitted data.
- This approach filters data similar to how particle physics experiments filter collision data from the LHC to identify rare events like the Higgs boson. Centralized cloud indexing and
3. The CLOUD: a foggy concept
More of marketing
buzzwords than well
defined technologies
People have different
ideas on the meaning
of these terms
In videosurveillance
the term CLOUD have
been too often abused
4. Advantages of a CLOUD based infrastructure
• No need of in-house computing infrastructure
• No need of skilled staff for installation, maintenance and troubleshooting
• Reliability. The provider takes care of:
– Redundancy
– Contuinuity of operations
– Backups
– Disaster recovery
5. Costs scalability of a CLOUD based architecture
• No initial investment
• Predictable costs
• Pay only what you use/need
• Grow or shrink computing power, storage and bandwidth on demand
8. Ever growing number number of cameras
Human surveillance is not affordable or feasible
9. A real waste of resources
Interesting events are just
a tiny part of the recorded video
10. HD CCTV = BIG DATA
• Data generated by video surveillance has grown to
practically unmanageable amounts
11. Real life case: July 2005 London bombings
The recordings were
examined for weeks by
human operators trying to
find a clue in the thousands
and thousands hours of
footage.
Weeks of recordings of
hundredths diverse CCTV
systems where collected
from diverse sources: city
center control, shops,
banks, etc.
14. Why cloud storage of video is not viable?
Sheer size of
produced data
Costs of storage
and transfer
Limited available
bandwidth
15. Moving mountains around...
… extremely rare, just like a speck of gold lost in tons
of rock
Miners don’t move mountains of rock around! They
bring mining equipment close to where gold ore is dug
Meaningful images are…
17. Dealing with Big Data
• Cannot rely on significantly more efficient image
compression algorithms
• Must rely on edge-side storage of high quality HD
video
• Must use video content analysis (VCA) to filter
important footage out (thumbnails or short clips)
• Describe meaningful events by means of effectively
searchable metadata
18. Extracting interesting information on site
Understand locally and communicate only if needed
Smart cameras can filter
out significant events to
reduce the quantity of data
streamed to the data
center even with limited
bandwidth
19. Blob Motion Tracking
Tracking & Trajectory
Smoke detection
Fire detection
Face detection
Crowd detection
Number Plate Recognition
Lost and Found Detection
Traffic Controls
Origin, Blind, Darkness Alarm Panic Detection
Available features:
The Smart Products
What can be done today
20. Accurate image
analysis:
now possible since it is
made on RAW images
coming from the
sensor
Smart Products
Internal Architecture
Local high resolution
storage and external
streaming with
adaptive bandwidth.
Broadband connection
not required
Bidirectional
communication layer,
encrypted and
automatic.
Secure and reliable
Internet connectivity
Virtualized data centre
software with Private
or Public Cloud
deployment.
The control room is
everywhere an
Internet connection is
available
21. Similar problem: the search for Higgs boson
CERN - LHC accelerator
• Bunches of protons and
antiprotons crossing 40 million
times per second generate about
20 collisions per crossing totaling
about 1 billion collisions per
second
• The frequency of producing a
Higgs boson is extremely rare:
once in 1013 = 10 000 000 000 000
interactions or one every 3 hours
ATLAS experiment
• ~100 million electronic channels
• If all data would be recorded, this
would fill 100 000 CDs per second,
a stack which could reach to the
moon and back twice each year.
• Online data filtering is then a must
• Level 1 trigger filters down to
about 75 000 events per second.
• Level 2 trigger reduces it to about
2 000 events per second.
• The Event Filter then selects for
permanent storage about 200
“interesting” events per second.
22. The correct approach
• Normalize and correlate
information from
heterogeneous sources
to extract meaningful
and actionable results
• Merge information from
widely distributed
different providers and
organizations
23. Architectural requirements
• Distributed uncompromised information
• Hierarchical online filtering
• Centralized database indexing with simple and efficient search
methods for multiple distributed tenants
• Extraction of readily actionable information
• Possibility to instruct the devices on the field to execute online
and offline queries based on requests issued by tenants in the
cloud
24. Database indexing in the cloud
• Thumbnails
– vehicles, license plates, persons, faces, etc.
• Metadata
– Number plates, face recognition, etc.
• Classification and profiling
– Person
• Age, Gender, Etnicity, Height
• Mood, Facial expressions
– Vehicle
• Car, Bike, Motorcycle, Bus, Van, RV, Truck
• Manufacturer, Model
25. Actionable feedback generation
• Correlation
– Timestamping (synchronized)
– Georeferentiation (Indoor, Outdoor)
– Identification
• RF, NFC, BLE, wifi, cellular, loyalty card
• Voice and Biometrics
– Third party databases
• Publicly accessible data
• Crowdsurced data
• Data accessible only by Governmental Entities
26. How to get there
• The investments and the time needed to achieve the described scenario are
certainly very large
• It is important, then, to devise effective ways to normalize, merge and
analyze data coming from existing systems preserving most of the prior
investments
How can some information be secure if it is available in the cloud
Probably, the CLOUD, today, is more of a marketing buzzword than a well defined technological concept
Asking ten different people what the cloud is, you are likely to get ten or, maybe, eleven different answers
In the field of electronic security the term has been often abused gaining an inappropriate meaning or, worse, a negative misconception
It is not affordable/practical to monitor a large number of cameras both for costs of involved resources and for human inability to cope with prolonged attention. According to research, the average professional operator attention drops to only 40% after only 30 minutes. In case of multiple sites spread in the territory local monitoring is too costly and centralized monitoring becomes a must.
In a security system the interesting events worth being looked at or saved are, on average, in the order of minutes per week of a few cameras
All the new HD video surveillance cameras installed worldwide in 2013 have produced 413 Petabytes (1 PB = 1000 TB) worth of data and are expected to produce more than double in just four years, expanding to 859 PB in 2017
The belief that the CLOUD is just something related to remote storage is widespread in the security market
It probably comes from the most common use of the cloud for storing large files, e.g.: File storage/sharing = CLOUD = DropBox
Centralized storage is in most cases unpractical and often impossible because of the sheer size of produced data and limitations in available bandwidth
The mentioned amount is enough to fill 92.1 million single-sided, single-layer DVDs or it’s four times the amount of photo and video data stored on Facebook as of February 2012.
Transferring and storing large amounts of video to the cloud or to a remote data center is economically unsustainable
The surveillance business is adopting several technologies designed to accommodate and mitigate the rising tide of data.
New data compression algorithms. For example, the High Efficiency Video Coding (HEVC) standard—also known as H.265—has been claimed to double the data compression ratio when compared to H.264.
Video content analysis (VCA) can be used to filter important events out rather than simply recording continuously.
Nevertheless recording of HD images will only be feasible locally moving at a central location only important images and metadata produced by VCA and biometrics
Only the capacity to normalize security data coming from heterogeneous sources and to correlate the gathered information to be compared against known patterns will allow to extract meaningful results which could be actionable in a timely manner
It should ideally be possible in widely distributed systems to merge information coming from different providers and organizations