A jobsite at Temple University bordering dense residential neighborhood had a temporary fence to protect deep excavation pit, but fence panels on uneven ground had up to 10” gaps below them to ground — plenty of room for a child to slip through. This is typical of a small detail that goes unobserved which can lead to multi-million dollar liability and tragedy.
By Ken Hantman, President, Perimeter Protective Systems
This article is intended for criminologists and security directors at hospitals, universities and other large campus-type facilities. This is the first-known quantification of situational awareness and the economic comparison of machine and human intelligence to the task.
I have a 35 year overview of security and a five year deep study of video analytics and a one year study of BRS’s AISight behavioral analytics software. Based on that I recognize the significance of the advance that AISight represents, created by about $100 million in R&D funded through Homeland Security and other investors. This paper is the first attempt I am aware of within the security industry to quantify to a first order of approximation the changes in situational awareness yielded by artificial intelligence.
All other video analytics are rule based. They are superior to AISight if the intention was perimeter protection or setting up defined virtual fences or defined “regions of interest” mainly concerning intrusion into such areas by people or vehicles during defined time periods. This has extremely limited application to a university campus because, simply, there are thousands of people who belong all over the campus at all times. AISight is the only practical analytic system that can function in such an active environment.
Rule-based analytics will not work there. It cannot distinguish between normal and abnormal or good and bad behavior. The only thing it can determine is “no one belongs here” or “anyone belongs here.” There are some other rules such as object removed or loitering time, but in reality it is “no one belongs here” that is utilized.
Only AISight can function to distinguish abnormal behavior and with proper monitoring to discern likely criminal activity or emergency response required situations. This is a great opportunity to be acted upon.
You must understand the nature of this software because it is not like a traditional burglar alarm or even like motion detection video. Defining what is of interest to you as an anomaly versus what the machine thinks is an anomaly is a process of refinement and to a degree thinking into the mind of the machine and having the computer-code-writing engineers translate or “tune” things. Hence the user is invited into a partnership as it were, the purpose of which is to bring a level of security far beyond what any other technology can offer. It converts a half million dollar extant infrastructure which provides no useful information in real time into a system which provides useful info, and does so at the negligible cost, relatively speaking.
AISight does not have the same range as rule-based analytics such as VideoIQ which can work reliably at 300’ with fairly wide, say a 28 degree angle-of-view lens setting. AISight with that angle of view will work to about 80’ out maximum for human activity, farther for bigger things like trucks or fires, and less far in non-ideal conditions. So if you have hi-def cameras and monitors you can presently see with your eyes a much bigger area than AISight will see for human activity. Accordingly you will not have complete coverage of all areas immediately. In the future as processing speed increases, the software will be able to process in real time more pixels and the effective range will increase.
Via traditional video cameras presently there is no situational awareness in any other than an almost infinitesimal way—randomly observing the screen.
I’m not aware of anyone having previously attempted a quantification of situational awareness but here is my product.
Summary
If we divide the acres of campus A (317) by the number of patrolling officers (about 25 per shift) we find there is about 4% situational awareness coverage for outdoor areas and significantly less for indoor areas. This is an extremely rough estimate. It is based on how much space a human officer can effectively perceive, related to angle of foveal (central) vision, attention, and particularly the numbers of obstructions to lines of sight.
Methods
The assumption is that half the officers are outdoors. We assume for the calculation that a person can see about a half-acre of space at a time, where there is sufficient resolution or acuity to perceive “situations” which would be things like someone grabbing someone, someone pointing a weapon, someone climbing on a statue, laying on the ground, and so forth. True foveal (line-of-sight) vision is only several degrees field-of-view but eyes constantly scan to the sides of this to gather information to the brain, so the computation is based on a 45 degree field of view. Because there are frequent obstructions such as trees, buildings, people, vehicles, we have assumed that on average a depth of field of 200’ is utilized.
So, 317 divided by 12.5 (half of the 25 officers on duty who are outdoors) = 25.36. This means each officer is covering 25.36 acres but can only really see 0.50 acres at a time. So, 0.50 divided by 25.36 = 0.0197, which is slightly under 2%. At any given moment there is only 2% visual coverage of the space in an instantaneous way.
Now, humans have many capabilities. An officer may hear a scream, almost immediately causing focused directional attention. Given that dangerous or aberrant behavior persists over time --- someone climbing a statue will be there for seconds or even minutes --- the human equivalent of pan/tilt/zoom, namely looking around, will effectively change the ½ acre awareness into 1 or even 2 acre coverage. This corresponds to converting the 45 degree field of view to 90 or even 180 degrees. But that is an absolute maximum because such vigilance is rare and hard to sustain. Mental factors of fatigue and distraction would reduce this. Physical factors of glare, dim light, and partial obscuration will reduce it and so will environmental factors of rain, snow or fog. All told, as a reasonable first order of approximation considering the factors discussed we may conclude that a median of 1 acre coverage is a fair assessment. This corresponds to 4% outdoors situational awareness.
Indoors the situation is much worse. Walls dividing rooms prevent lines of sight. However, identification of strangers is more readily accomplished in this area. If staff or students have means to communicate to Security then in the more controlled indoor environment the need for situational awareness may be less than outdoors. However, in isolated hallways or rooms or areas the risks still exist.
Entirely dependent on the location and coverage of the existing cameras, AISight may increase this situational awareness from 4% existing up to 50%, which would be a 1250 % improvement. In some respects this is the equivalent of hiring an additional 143 security officers per shift on patrol. That would cost $1.3 million per month extra, as opposed to approximately $7500 per month extra. (Costs are based on a long term lease of the software or short term amortization of capital purchase of perpetual license; costs of hardware or servers is not excluded from these calculations.) (Naturally if there are no cameras covering certain areas and within about 80’ of the targets, then results will not ascend as spectacularly. Further, by the nature of AISight we don’t expect more than 30% observation even within the camera covered areas and ranges, so we should probably say it might improve awareness by as much as $390,000 per month, in this example.)
So, as to cost and efficiency we can conclude that AISight is in excess of 50 times more cost-effective for situational awareness than humans. ($390,000 divided by $7500 = 52). (A similar calculation was done for University B and in that case AISight was in excess of 20 times more cost-effective than humans. The difference relates to how many acres of campus, how many patrolling officers, and how many cameras).
There is little surprise here that the machine beat the human in increasing situational awareness. Calculators are thousands of times more effective than humans doing arithmetic. Computers are tens of thousands or even millions of times more effective than humans in storing information or correlating it in some ways.
Of course a live human can do many things which a remote guard cannot do. Also visible physical presence of officers deters crime. Because seconds matter, the best use of AISight would include adding audio-out capability to cameras if it doesn’t presently exist. This would involve a separate hardware cost. Assuming equal dispersion of officers on the campus map, if half of the 25 on shift are outside, then they will each be on average 250’ from the locus of an incident when it happens. If responding security officers on bicycle, and certainly if on foot, this will take much longer for them to get there than the remote audible talk down would take as an immediate first response step.
Consequences of Improving Situational Awareness
By bad luck a campus tragedy is a spin of the statistical distribution roulette wheel. Then the TV stations’ cameras and microphones are thrust into the president’s face. Perhaps unfairly, but in the emotionality of the situation, they look at the security director and say why did this happen? You can try to explain there was no way, no technology to prevent it and unless the university had budgeted for 1850 officers they could not possibly observe all the campus (and even at that number could not observe in each room in each building). But AISight does change the equation at least significantly. So the reporters’ questions would become “Why didn’t you get this for $7500 per month?” As an expert witness for a plaintiff I would certainly make that case. Indeed an employer or institution may be found to have a duty to its employees, students and visitors to protect them and to reasonably avail itself of available technology.
A new PSIM, which will cost hundreds of thousands, is certainly a good thing. But it is doubtful it will achieve actual improvements in security in any way equal to AISight. It is a good thing and the associated sales pitch claims that ultimately efficiencies achieved will pay for the PSIM over time. But it is also quite a stand-alone thing from AISight. The cost of AISight is trivial compared to the PSIM, while the actual security advantage of the PSIM is probably trivial compared to AISight.
Although security is best when it contemplates things that have never happened, it is also useful to consider things that have been prevalent. Drunken behavior is a typical campus problem. Does it manifest in a way that AISight might observe? Do students who are drunk act differently than those who are not? Do they climb up in weird places? Fall down? Wander or stagger around? Lay on the ground at odd times of the day or night or season? Do crowds of drunken students act differently? Are there cameras in areas where these things occur?
Reviewing most Security and Fire Safety Reports it is clear that universities are a pretty safe place to be, the headline news notwithstanding. Tragedies happen but less frequently per population per square mile in the surrounding communities. But the nature of randomly distributed crime is that this picture could change with one tragic incident. While we cannot of course assure, and we specifically do not assure, complete protection which is obviously impossible, you can act to make a real improvement in situational awareness. All or nothing is not the reasonable philosophy and if something can be improved 30% compared to 0% that is a tremendous gain.