Last month, I wrote about big data analytics and its ramifications for security from the perspectives of information security and crime event correlation. This month, let’s examine this topic in light of its potential impact on video surveillance.
To recap, Big Data is a name which has been given to masses of data that have been, or are continuously, assembled from an expanding variety of information sources. These sources may be internal or external to an organization; produce data that is structured, unstructured or random; and which heretofore have defied conventional analysis. Now, companies such as IBM and RSA (division of EMC) are developing means to derive underlying value from these stores of data by surfacing trends, patterns and apparent correlations.
Video presents special challenges. At a recent conference, I heard a 2011 statistic, that 220 million frames of video are recorded every second. Assuming an average recording rate of 10 fps, that’s 22 million cameras feeding storage media, which does not seem unreasonable. Another 2011 study estimates nearly 2 million cameras just in the U.K., or about one for every 14 people. Given that IP CCTV cameras are outselling analog cameras, and that network bandwidth and storage capacity are increasing while costs are decreasing, it is safe to say that the amount of video stored per month will continue to increase — although it is not known how much of that will be archived beyond 30 days.
This point is significant — if video is not kept for more than 30 days, there is a finite time window for analysis; thus, if big data analytics for video becomes a game changer, that in itself may argue for longer retention periods.
So where is the opportunity? I believe the answer is in the correlation of captured images and sequences with similar material or with other data that is accessible by the system. We have an example of that (though not big data) in the retail market, where video is correlated with POS (point of sale) transactions. In this simple example, a transaction “tag” is placed on the video, enabling an instant view of what transpired at time of transaction. Other types of tags are time/date, alarm inputs and video analytic-related.
Scaling up to use “big data” to find events or to evolve relationships that are not obvious involves several requirements for the technology to work as intended. First, video will need to be marked at the edge, or at least before it hits the storage device. Marking at the edge can be based on the original video data, derived before compression or frame rate or resolution reduction. This marking would include metadata, data or characteristics associated with the video image. The metadata is then transmitted and retained along with the video. Today’s IP camera and video analytic technology is just at the threshold of this type of capability.
The more metadata that can be generated, the better the chances that some correlation can be found by a product mining that data, which would involve the application of multiple analytic algorithms on a single frame of video. That could include pre-established rules or a behavior-based algorithm, such as that from BRS Labs (Request more info at www.securityinfowatch.com/10482504), which learns how to spot abnormalities in a scene.
Further big data mining opportunities exist in the use of facial recognition. Currently, facial recognition requires a high-quality megapixel (1.3 MP or better) camera and lens. In addition to the facial image capture, metadata about that face is stored and used as the basis for further comparison, such as the distance between the eyes. Add to this data information about clothing color, height, body shape, etc., and you have a potential people classifier that can be used for further analysis and correlation. Examples of this could include immediate identification of a person of interest in a crowd or searching a segment of stored video for all instances of people matching a target profile. Some would argue that that is possible now, but big data implementations will dwarf the scale of today’s limited systems.
Edge marking can mean within the camera, but the intelligence may also be contained in an intermediate processor or within the network equipment. An example is Cisco’s Unified Computing System (UCS), which combines network transmission with high density computing and is capable of hosting analytic applications that can be applied to incoming video streams (request more info on Cisco at www.securityinfowatch.com/10482443).
Between smart cameras and these types of smart networks, we can surely expect to see video stored in a way that makes it far more useful than most stored video is today. Still, the value of all of this will likely be more in the forensics area than in traditional security. That said, big data analytics for video has great potential in enhancing the value of the security infrastructure by creating added business value, with operations efficiency, customer satisfaction, liability reduction and marketing data as areas ripe for further development.
It all comes down to making sense of a ton of information that, right now, is of limited use.
Ray Coulombe is Founder and Managing Director of SecuritySpecifiers.com, enabling interaction with specifiers in the physical security and ITS markets; and Principal Consultant for Gilwell Technology Services. Ray can be reached at [email protected], through LinkedIn at www.linkedin.com/in/raycoulombe or followed on Twitter @RayCoulombe.
