Video Analytics- Unbiased Articles/Papers

Hi guys and girls,

I have designed a system for a client using video analytics as a method of alerting if fencelines are breached (line crossing) or people linger in certain zones/areas (intrusion dection). I will be using Hikvision Darkfighter cameras- well featured, priced right and perfrom very well. The client has an existing VMS- Milestone Professional 2014. As Milestone does not (currently) have a method of utilising the analytics of the camera directly, I was planning on setting up the analytics within camera software, conecting the cameras output to its input, and then using Milestone to provide the alert when it sees the cameras input change state. Not ideal I know, but under the current Milestone software, I see no other way.

I can cover the required fencelines with CCTV, so placement is not an issue, nor networking or lighting etc.

But this is not my current 'problem'.

What my problem is, that another company is also designing/quoting. They are proposing a physcial beam set up (many of them) with fence wire detection (wire in the fence picking up the netting being cut etc).

My reasoning for not using beams or fence wires is:

The site places containers beside the fence- beams would then be broken, causing false alarms or being turned off. Items being stacked, providing something to climb over- espcially as a neighbour legitimately places items against their side of the fence. Same reason wht a fence wire won't be reliable- simply climb over etc. As well as needing to keep beams clean, easy to get round, physical damage (hit by trucks, forklifts). Using a line crossing anaylitcs still will cover the fence line even if ojects are placed against the fence.

However- as the other company has (unprofessionally) bagged our design, the client has asked for any documentation (independant preferably) proving analytics (line crossing) works.

We also have an access control component, so having to prove our widely used, full reporting, multisite system compared to the other companies small, one site, antiquated system, is better also.

All this after an extensive proffesional 2 hour presentation of a working system....

The site will have a mix of camera locations (general areas overviews, ID on 6 entry points, general internal overview (processing/factory areas) as well as our fence line cameras. Totaling 75 cameras on site.

Anyhow, sorry for the detail, but does anyone have (or links to) any doc's or papers on anaytlics that shows them in a positive light?

cheers,

Glynn


Glynn,

Before offering any feedback, I just want to make sure I understand the comparison correctly. You are looking to use Hikvision's camera built-in analytics and want evidence / validation that they perform equal or better than physical beam / fence wire detection. Is that accurate?

Essentially, yes. Does not have to be specific to Hikvision, but require evidence/validation of performance of analytics.

We do have sites using line crossing for people counting at malls, but this apparently is not enough.

thanks

Glynn,

There's a lot of variation among analytic providers and analytics setups so it is not sensible to provide a 'general' validation of analytic performance. Analytics is not a mature enough technology / category that one could rightfully prove across the board performance.

I apologize for Undisclosed A's comments, the tone is unprofessional. However, some of the criticisms / concerns alluded to in them are valid.

In terms of unbiased articles and papers, IPVM has them but they point to a much more problematic conclusion / outlook than what you are looking for. Beyond that, no credible source would make categorical positive statements about analytics across manufacturers.

Specifically, have you validated how well the Hikvision cameras / analytics work on that site over time?

LOL

excuse me? whats so funny?

Okay, maybe I started celebrating the holiday early. I have a background in analytics and using it indoors is completely different than outdoors. Every analytic operates differently so proof or evidence of one version working to validate another just isn't going to work. Using traditional CCTV cameras in a perimeter application with video analytics hasn't proven to be very successful although some will claim accuracy. Are any of the cameras facing east/west? Are there any streets running perpendicular to the fence line? (Headlights) Are there any streets within about 100' of the perimeter (emergency vehicle lights) I agree you are in a better path with detections for loitering and such a beam system will never detect. I just can't tell you using a competitor's evidence is accurate and each application is different. Review the many articles on analytics IPVM has done.

Headlights nor emergancy vehcile lights will not be an issue- no fenceline we are covering is street frontage, or can be influenced by such lighting/change of.

Camera positioning has been carefully calculated to minimise sun inpact.

Cameras are positioned to look at a fence, with the line crossing running across the image, rather than 'down' the fenceline- giving a more even overall distance. Testing has been done and has proven to be better performing than expected- to the point where I am very confident in using it.

If the line crossing is breached, an alert is sent to the monitoring station. They bring up the camera image and verify the trigger. If required, they then can speak to the person over a loudspeaker.

We match the triggers in to our access/intruder system, giving us the ability to bring up associated footage off multiple cameras at the tigger time.

Sounds very well thought out. Let me know how it works when the season changes.
This is in new zealand. We don't have a extreme season changes as some other parts of the world. While the sun will be lower in winter, it won't be shining in to the camera causing issues. Unless I am missing something else?
New Zealand. In the immortal words of Napoleon Dynamite "Lucky!" Exterior is unpredictable. You will have to manage an expectation between Probability of Detection (customer wants 100%) usually and False Alarms (customer wants 0%) typically. Going into a central station is certainly a great idea and helps. However I met with a person at ISC who felt the 2500 alarms per month he was getting with his current installer was excessive. That came from the central station. You stated that during design and testing you identified the risk of a blinded camera and minimized it. I didn't read "eliminated". I read minimized. So how many hours is the customer OK with limited or no detection? In case you are wondering I think the beams will be worse

Dear Glynn, I was informed my comments could be taken as a personal attack towards you and as I read them again it's easy to see. That was never my intent, my sense of humor is definitely on the edge and I apologize. You asked for an unbiased opinion and actually I have a bias since I am very familiar with a analytics. What I should have said more clearly is the each analytic provider filters / analyzes and processes using a variety of methods. Trip wire is usually the least accurate because it rarely classifies as human and then alarms based on location. Typically trip wire is just a group of pixels crossing an area. Some advanced motion detection defines "must be taller than wider". Some analytics will use a form of intelligence where they have pre-loaded the shapes of a human or vehicle from many angles to compare. True analytics will have less also alarms in an outdoor environment and will catch things like a "half human" when in a wheelchair, riding a bike, etc. So unused words like "typically, usually, likely" to describe these as each product is different. I was casual and it was an attempt (poor as it was) to caution you that video analytics has chewed up many a budget with configuration and return visits. Your ideas, design and efforts in advance should be applauded and my only fault I can see would be the over marketing on each of the manufacturers. Some work better that others. Be careful in what you promise and manage the expectation. Obviously the other guy isn't but I doubt you can fix that. The most serious comment I did make was a passion for New Zealand. You have an amazingly beautiful country and there was no jest in that statement. I always tell my kids the truth comes after the word "but" so I'll leave you with my heartfelt apology.

Thank you.

To all that may be following my request:

I am not expecting a 100% perfect operation- neither does the client. I do expect though it will be more reliable than a beam system. Ideally, I would like some form of document that might show the comparison between the two.

The cameras which will be used for the line crossing are looking at the fence - giving a fairly even sensitivly across the line.

There is very little chance any object the size of a person, who is not a person, that will cross the line.

If anyone one has links or information that may be useful, please let me know.

Thanks,

Glynn

"Ideally, I would like some form of document that might show the comparison between the two."

Outside of a manufacturer biased post, I do not think there is anything out there actually doing a study between the two.

"There is very little chance any object the size of a person, who is not a person, that will cross the line."

The problem here is that many video analytic systems mistake things that are not the size of a person as being a person. It does not sound sensible but even the best computers / cameras are nowhere close as being as intelligent as a child in recognizing persons. Whether it's wind, shadows, rain, snow, wild animals, low-light noise, etc., it only takes one issue to cause a flood of ongoing false alerts.

If you have tested those Hikvision cameras, with those Hikvision analytics in the actual positions the deployment will be and have verified that total project false alerts are under acceptable levels, great.

The reason we are so cautious here is that we have seen a lot of integrators win significant analytic projects and then have them blow up in their faces, losing money on the deal and having very frustrated customers.

"The cameras which will be used for the line crossing are looking at the fence - giving a fairly even sensitivly across the line."

One other thing, most analytics are worse detecting activity coming towards the camera then moving across it. Again, if you have done extensive field testing validating this, you are safe.

If you have any test clips from the Hikvision analytics I would love to see them.