What People Counting Analytics Should I Use?

I am a Security Technology Consultant based in Perth Australia. I am currently involved in the design and specification for IP CCTV systems for three shopping centres, all of which wish to use video analytics for people counting at entries.

I have looked at three solutions so far:

  • Agent VI interfaced with OnSSI’s Ocularis.
  • SenCount by Sensennetworks (VMS independent)
  • Cognimatics TrueView (VMS independent)

I would like to know of anyone’s experience in this area and their recommendations. I have only seen one operational system so far and that is the Agent VI with Ocularis. There is a system in a retail centre in the Perth CBD monitoring five mall entries and this system appears to be reliable and accurate. It utilises cameras set up to “look down” directly over the mall entries. These cameras are not used for general surveillance and are specifically assigned for people counting. The Centre Manager runs reports on a monthly basis of shopper/pedestrian traffic throughout the mall areas. This information is used by their leasing agents to value add the leasing of areas.

Sensenetworks state that SenCount can use angled view cameras that are also used for general surveillance but I believe some conditions apply. The system is installed in a large University in Melbourne and to count people using the Sydney Ferries.

Cognimatics software is embedded into AXIS cameras that are specifically assigned as people counters in “look down” view. This is the only solution out of the three that does not require a dedicated server. Haven’t found a live application yet.

I look forward to your comments.


John sent an e-mail asking me if I can respond to the discussion you have initiated. I am cutting and pasting the response I sent you when you contacted me.
"You are right in that we can deal with a wide variety of angles - typically found in CCTV installations without compromising accuracy. We also have specific customisations to ensure we deliver the accuracy that is demanded by the customers.
We look at the accuracy as follows: (Total number of people manually counted - total counted by system)/total number manually counted.
Our counting accuracy in all practical installations can reach in excess of 95% and often reaches 99%.
We have successful implemetation of people counting using this approach including Abu Dhabi Airport, Manly City Council, Colonial First State managed buildings like Roxburgh Park and Broadmeadows shopping centers, Queen Victoria building in Melbourne Australia etc. Details of two of our successful implementations include
  1. People Counting on Sydney Ferries (All ferries used by Sydney Ferries now have our people counting solution) - Accuracy was verified by the end user by employing people to count by clickers at the entrance of the ferry and verifying the counts produced by cameras. We delivered >96% counting accuracy in all weather conditions 24/7.
  2. People counting at Swinburne University of Technology (Library) - The accuracy is verified to be 98% by looking at the recorded video and manually counting and comparing with the counts produced by the system
All our deployments use existing CCTV installations and perspective views. Also, statistics 101 suggest that we should use a minimum sample size to measure accuracy - in this case a minimum of 100 people should be used for in and out accuracy."
The additional information I wish to add to what I wrote earlier to Robert is that we now have an embedded solution that works inside the Axis cameras based on latest Artpec 4 chip and does not require a dedicated server.
Subhash Challa
CEO - SenSen Networks

Subhash, can you elaborate on what angles you are claiming to deliver people counting? 15 degrees? 30 degrees? more? Can you share an image of what the steepest angle would be?

What rationale do you have for being able to deliver this accurately at angles when seemingly none of the other major players claim to do so?

John,

We can ensure accuracies in high 90's up to 45 degrees. We have gone higher than 45 with some drop in accuracy and in some extreme cases we have to customize a solution using other analytics available in our platform. For example - in some situations when the ceiling heights are short, we delivered people counting using detected Faces. This kind of customization is rare and often not required. Here is a snapshot of some of the angles and conditions from our customer sites.

Most vendors require top down view as their techniques require clear separation of objects. Our technology is different and it does not require full separation of objects to work accurately. Most vendors also have height restrictions and associated lens/focal length restrictions. So they seem to have standardized their solution by specifying the heights and focal lengths of the cameras and look top down to ensure separation of objects.

On the other hand our solution works with most angles encountered in Surveillance camera deployments and does not have any height restrictions per ser as the technology is completely different. In our approach, to ensure accuracy, we need to calibrate each camera view for accuracy. We handle this by requesting the integrators to send us the video from the existing views and we send them back calibrated settings files - that ensures accurate counting for those views.

Subhash CEO SenSen Networks

Robert,

Aimetis has integrated people counting as part of our VMS. We also have embedded people counting application running on Axis network cameras (similar to Cognimatics). I've recently been to Perth for a project actually. Please contact me to discuss further if interested.

Justin, can you comment on how Aimetis does people counting? angles? reporting? accuracy? etc.

I have used the Aimetis for a couple years now to count people for building capacity. Ill say its as accurate as a low paid employee counting with a clicker. It always within 10 0r 20 people of what they count manually. That said its pointed straight down at the proper height in a choke point area that our guests have to go thru to enter and exit. Lighting is also a factor for it.

I had planned to change it out for Cognimatics TrueView last spring so I could use multiple cameras and thier web reporting software but got too busy. But from what i tested it was a little better even than the Aimetis.

Thanks for your comments John. You have mentioned it is within 10 - 20 people accuracy but out of how many people (percentage accuracy)? I have been trying to arrange a demonstration of Cognimatic's TrueCount in Perth but Cognimatics wont provide their distributor with a demonstration copy of their web reporting software. This is truely frustrating as I have real opportunities and they may miss out because of this. I believe in "try before you buy".

I am flying to Melbourne to take a look at SenSen Networks SensCount and one or two of their reference sites mentioned in Subhash's message. I will give an opinion upon my return. This solution may fit well with two of our current projects where it is impossible to install cameras looking down directly over the entrances.

Hi Robert,

We distribute Cognimatics in Baltic countries, Europe, I have more than 5 years expierence with their products. As you noted it is edge analytics embedded to standard Axis ACAP cameras, so you don't need any other devices on site. Cameras must be installed vertically above counting zone to reach high accuracy, which usually is 95 and more percents even in high density flows, we have counters on spots where more than 10.000 passages per day are counted. You can sync several cameras to each other and built virtual counter for any zone width. Statistical data are pushed from cameras to web server, hosted by customer, system integrator or Cognimatics, where variuos reports can be created by unlimited number of users. System is really accurate, stable and easy manageable.

I don't know who represents Cognimatics in Australia, but if you need to see live system, test demo licenses or account to Web report - just let me know.

Box&Value, distributor of Cognimatics

I'm pretty sure Cognimatics will give your distributor an Eval code, in fact they gave us one as a Reseller.

I would suggest that the Cognimatics solution would produce stronger accuracy than any solution using a generic counter. If you want really high accuracy, you should consider a Brickstream, again a dedicated overhead counter camera with high accuracy including high people density environments.

Hi Robert. I wanted to inform you of Verint's people counting solution - part of our video business intelligence platform. Our solution helps users make informed decisions by providing people counts with an accuracy rate of more than 95% – which is higher than what traditional technologies can produce.

Courtney, can people get Verint's people counting solution for use with other VMSes or is it tied to Nextiva?

Hi John. The Verint People Counting solution does not require the use of Nextiva Video Management Software, although that is how it is typically deployed. We are an open platform and can work with other VMS' as well.

With respect to several comments in this thread, I'd like to clarify a few items regarding Agent Vi:

- Our counting analytic can be run on cameras that are not vertically oriented. Other analytic rules can be run on these cameras to benefit from the increased field of view. I won't challenge Subhash's claims since I don't have first hand experience with their system. However, there is always an inherent trade-off for video analytics related to field of view. (EX. A camera that can't see a child standing next to their parent due to the look angle. VA can't analyze what it can't see clearly.)

- Agent Vi is an open system that works with many VMS providers. We have an excellent integration with Onssi as well as Genetec and Milestone just to name a few. We are VMS agnostic.

- Our system is edge + server archictecture. Our agent runs in all major Axis ACAP enabled cameras. The heavy video processing is done in the camera and the server performs the rules analysis as well as collecting business intelligence. The server usage is light and doesn't need to be on a dedicated server. This allows us to perform analytics and forensic search across multiple cameras which is a big value to many users.

Thanks

You are absolutely right - one cannot detect or analyse what the camera cannot see. That said, most surveillance camera installations by comeptent SI's do ensure that all of the objects that they want to monitor come into view atleast partially. Once such views are available to our SenCOUNT solution, it will make the best use of that information to ensure that the count estimates produced are as accurate as possible and go a long way in meeting the specifications of end users.

As I clarified earlier in this chain, depending on the scope of project and the constraints, we have a solution that works with Axis latest Atrpec 4 chip based cameras with processing done on the camera and a light weight server application on the any machine to support business intelligence. In some cases where various different methods may have to deployed - a sever based solution is specified. The ultimate goal is to deliver the most accurate counts irrespective of the site specific challenges and requirements for end users.

Hope this clarifies.

Subhash

hmm... this is a good topic. Anyway to get IPVM to "test" them out for comparison?

Margarita, I'd love for us to do this at some point in the future, but cameras come first and we still need to get going on new VMS testing. We are working on expanding our team but finding the right people in the right places takes time.

Thanks John, I understand!

It has certainly been an interesting exercise! It appears that there are many solutions available but few that can utilise angled camera views. We have seen SenCOUT in operation at Monash University library (Melbourne) and yes it appears to deliver accuracy to the levels stated by Sensennetworks. The client is very pleased with the system and the support offered so this is a big tick to Sensennetworks. We now have Agent-VI stating they can achieve simialr results so JVSG representations of camera views have been supplied for review.

What will be important in the evalutaion of the solutions offered is the support offered to the Security Integrators during the installation and commissioning phase and the on-going support.

It appears the market, particularly retail see real ROI on an investment in this technology so I believe it is a growing trend. Maybe it is something IPVM can put on the roadmap for future evaluation?

It's certainly worth IPVM testing in the future. Robert, thanks for bringing this up and for the feedback.

I have jsvg & love it .

Question : How does this work with ascending , decending targets with multiple sizes arranged together .

Large person next to small person or small persons going down escalator in front of large person.

Then when you cant funnel into sequential order how does this work ? How accurate ?

Large Quanities of people , concentrations of mass ( terminals , inlets , outlets )

Just out of the norm or controlled enviornment show here .

Dear Christopher,

I take the opportunity to respond to your comments about large quantities of people or concentration of mass. Evitech has released LYNX, a crowd management software addressing these specific concerns.

Information and White paper could be found here:

Peaceful Management of Crowds + whitepaper

Is it what you are referring to before I comment further?

Christopher - Responses Below:

Question : How does this work with ascending , decending targets with multiple sizes arranged together.

Large person next to small person or small persons going down escalator in front of large person.

Then when you cant funnel into sequential order how does this work? How accurate?

Subhash: While asending or desending they all have to cross a counting line somewhere in the scene. Counting lines need to be chosen that maximize the visibility of people of all sizes - weather arranged together or not. Once visibility is maximized, calibrations done on selected videos will take those variations into consideration in providing count estimates. Accuracy in most situations will be greater than 95%. No need to funnel people sequentially.

Large Quanities of people , concentrations of mass ( terminals , inlets , outlets )

Subhash: Top two images in the first column is from Manly Beach in Sydney - It attracts huge crowds over weekends and some festive times. I will dig up images from the live site showing large crowds for your reference.

Just out of the norm or controlled enviornment show here .

Subhash: There is nothing out of norm or controlled there. They are actually from our live installations. The images are from indoor and outdor locations. The bottom row is from our live installation of Sydney Ferries - They are moving, wobbling ferries where people counting has been successfully implemented. Happy to share more data and customer contacts for your own first hand verification. Other images are from libraries and shopping plazas. Manly beach had Mobotix cameras already pre-installed before taking our SenCOUNT solution. Sydney Ferries already had Axis Cameras before they introduced people counting solution. The shopping center had analog cameras pre-installed before taking SenCOUNT solution.

Hope this clarifies. Please do contact me on subhash@sensennetworks.com if you want further detailed analysis and information in this regard. I cannot share customer/partner contacts in this forum.

Subhash

Subhash

Excellent Reply , pdf explains in detail the concepts & reviews

Thanks

Robert

Video analytics has two basic platforms. The first is the 'edge processing' camera, including monocular and the more sophisticated stereo vision cameras, which capture an object (top-down) and translate the activity to data, which minimize broadband requirements and also allows for higher capture rate and more demanding counting requirements such as queue and in-store analytics. The other system analyzes images captured by a camera on an in-store/mall server. The advantage here is cost and the ability to use the cameras for security. The challenges are in accuracy. Currently, for a relatively simple door-counting, both systems and most technologies are 'good enough'. To choose a vendor, check out not only the system but also the requirements for calibration and the ability to support you.

The requirements for traffic counters include the tolerance for sunlight and shadows, temperature changes, high volume traffic, and a minimum of 95% consistent accuracy. Best practices of auditing include volume, granularity and consistency. As mentioned, the minimum volume of tested traffic is 100. Granularity of the audit depends on the application, and shopping malls generally don't require less than a day, but retail stores require hourly or 15 minutes segments. The test of consistency is crucial. If the system both over-counts and under-counts, the counts masks the nature of the error and in many cases renders the analytics less relevant. And for video analytics, do not audit against manual counting (after 8 years in the people counting business trust me on this), only against video clips or streaming video.

While most shopping centers do not require sophisticated analytical application, there is an interesting trend to compare between the number of people passing by (measured by the shopping center counters at the aisle) and people entering the store (measured by the store's counter, inside the lease line).This Capture Rate part of the general trend to monitor activities of the physical store as closely as retailers measure clicks online