Everywhere we hear that data is getting bigger and we’re getting more of it. One of the single biggest drivers of data growth is video. Video is becoming more pervasive, nearly everybody is carrying a video camera around in their pocket – SmartPhones. London now has over 100,000 CCTV cameras. The Metropolitan Police have issued over 25,000 body worn cameras to officers. In terms of data size, what does this mean? A one-hour police interview typed into a witness statement is less than 0.5 Mbytes. As a video it’s around 1,000 Mbytes – so 2,000 times the size, to ship around networks, to process and ultimately to store.
The normal way of working is for fixed CCTV cameras to send the video stream back to a monitoring centre and have people sat looking at it. Problems arise when there are many cameras or if the operator is distracted and misses something. There’s also the requirement to look back through historical footage because we have new information – we might now know that we’re looking for an individual in a red baseball cap, so want to see all of those candidates.
For several years we've been able to build central server farms that can “look at the footage” alongside operators and effectively tag the video with text descriptions of what’s in the video (this is known as ‘metadata’), allowing subsequent searching of the video - "show me all the people wearing a blue shirt, who are bald with a beard, wearing glasses that crossed the field of view from left to right". However, a fairly expensive computer (£75k) processing this sort of footage can usually only deal with around 20 hours of video per hour, or 20 simultaneous streams AND of course, you need plenty of bandwidth to get the video back to a central point in the first place.
Solving this sort of problem in future will certainly be done by running the analytics at the edge, or where the camera is. The processing will probably take place inside the camera, but right now, we don’t have battery powered processors fast enough to build into body worn cameras, able to do any kind of sophisticated analytics. However, for CCTV or fixed cameras this edge computing is becoming a reality today.
We've now built edge systems, where the video analytics is executed next to the camera (£1k piece of technology on top of the price of the installation and existing equipment). This brings a couple of immediate benefits. Firstly, if bandwidth is limited or costly, then we no longer need to send all the video, instead we can just send the metadata. Secondly, we can now program the camera to provide automated alerts. This could range from “alert me if a large vehicle enters the scene” or “alert me if somebody leaves a bag unattended” to “alert me if there is a fire visible”. One operator could now be effectively monitoring many more cameras. The other massive benefit of this approach is that it allows much simpler and more accurate searching of the video after the event. Now it’s quick for a search across all the metadata from all the cameras to find the bald person in the blue shirt. We can then go back to the edge device and only send the still frames of video that match the query.
This is a topic that is evolving quite fast, but much of the underlying software technology is now fairly mature – we’ve been waiting for hardware to become small/fast/cheap/low power enough to deploy these solutions and things will rapidly get better.
If you’d like to know more about how you could take advantage of this technology, please feel free to contact me on LinkedIn (http://linkedin.com/in/markgoossens) or Twitter (@MarkTheGoose).
Mark Goossens has been on the techUK Justices and Emergency Services Committee for the last 6 years and leads on Interoperability for them. He is also the IBM Client Director for the Home Office & Police.