Struggling With Video Analytics On Hikvision NVRs - Member Email

I have found that the line crossing detection is fairly accurate, but obviously not 100%, you will still get some false alerts but way less than you would if you simply just did all out regular motion detection. Intrusion detection is not bad either, but is less accurate than line crossing. I find a combination of regular motion and line crossing is the best route to go in certain scenes. I really like Line Crossing alot though, I am surprised to hear you say that it gave you so many false alerts.

Face detection, meh, it works sometimes but not all the time. I wouldn't recommend it for crucial scenarios. I also have to admit that Im not impressed with all the bleeding edge analytics that are coming out right now on the standalone units. I agree that we still have a little bit to go before they become accurate. We are working on a new OEM project with a new manufacturer that has alot of the "intelligent" features built into the camera. I told them to simply remove all those features because I feel they do not meet customer expectations.

I was very impressed with IPVM's review of Camio. I dont get impressed much anymore with new technology since it usually always lets me down, but that looked pretty nice.

You could add a conventional outdoor motion sensor, or use a thermal camera for detection.  Neither are cheap or provide a nice image, but they are “detection devices”.  

I've found that most of the VCA tools available with the entry level NVR's are pretty useless at minimising false alerts.

The intrusion detection can work well once configured well over a period of trial and error (percentage, time etc) however this can still trigger off false alerts.

I'm usually honest with the customers from the start, intrusion alerts are a great gimmick - but the false alerts may become more annoying than they are useful.

Video analytics have to synchronize with access control to set up multiple rules and applications. Think of a tensorflow (vms analytics solution) combined with navizon(database integration with access control database) applications.

Rule 01: Camera motion + Face Recognition --> Triangulation query + navizon ID.

Authentication = Face and Cell phone are true + Person has access to this area = No alarm. Metadata is still journaled. 

Authentication Failure = Person Face not identified or Cell Phone not in system = Instant Alarm. Navizon + VMS now live track individual.

Person's unidentified are broadcast to an SOC, email with photo, or to the nearest office or supervisor in the area.

Just a thought, do you have more rules you can think of?

Test Avigilon. They do object classification of people and vehicles first and then rules are created to provide alerts for various violations. If clouds (etc) do not get classified as a person then the rule is ignored and there is no false alarm. If set up properly (field of view, perspective, etc) there is high probability of detection and very low nuisance alarm rates. Nonetheless, expectations must be clearly understood and applied. Mis-application of any technology is a recipe for failure. 

Full disclosure, I am an Avigilon advocate. 

Sean, a couple of questions for our mate member:

1) How much could it cost to your Customer if a real intrusion detection is missed? More or less than 3.000$?

2) How much could it cost to your Customer to manage and eventually react to any false alarm occurred, for example in a whole year timeframe? More or less than 3.000$?

3) Do you have the possibility to export and provide some videos from those cameras where the 3.000$ NVR is definitely generating false alarms, just in order to be able to verify practically if other products can perform better? Of course possibly also some with real intrusions as well (just to filter all false alarms is very easy actually....;))..).

By the way, with less than 3.000$ you can surely buy a good enough PC with 8 licenses of some software products born and made for video analysis... But point 3 above would surely help me to give to you the assurance of this, or not....;))

Cheers,

Simone

I would ask the manufacturer (hik in this case) for guidance when it comes to using the analytics.  Surely they must have done a ton of testing and through all the trials internally and customer feedback they will tell how best when and when not to use them.  ie. how far to mount, how high to mount, what scenarios work best, etc....

Based on all the “unhelpful” marks, it appears a HIK fan was offended by comments that using visible cameras in the situation of uncontrolled external light sources is a bad idea.

Use a thermal camera.  You could use anyone’s thermal camera, including HIK’s and it will eliminate the effects of external sources outside of the desired detection area and allow the analytics to function better.   You will need a second visible camera to provide detail if desired.  Argue that. 

Use an externsl outdoor motion detector.  This will alert you based on motion characteristics and typically generate few false alarms bless there is another environmental hazard like water flowing.  This solves the lights issue. 

Ive sold and installed analytics costing thousands per channel and external lighting from helicopters, headlights and emergency vehicle strobes are very difficult for analytics to filter. 

Following up with links/references for the original post:

To add some detail, the referenced DeepInMind article is here: IPVM DeepInMind Tested Terribly

We have also tested other low-cost VMD/analytics: Dahua Intrusion Analytics Test and Hikvision VMD and Intrusion Analytics Tested

In general, we have found that these (and other) analytics perform reliably indoors but struggle in the outdoor conditions the member was referencing.