Using Face Detection On Samsung Camera

I have been using my Samsung SNV-6084R camera for a few months now, and have enjoyed great reliability with the motion detection. The camera is setup outside under an eave, and is set to the "Outdoor" presets. I do have a couple questions for the forum - hopefully someone has the answers:

1. Camera periodically "refreshes" its exposure: Periodically throughout the day, the camera seems to "refresh" its exposure, making the camera view either get much brighter (over exposed) or darker (under exposed), and then settle back to the correct exposure. It only takes a few seconds, and it's not a huge deal, other than the fact that this causes the motion detection to trip and thus causes a bunch of needless recordings...and it does this many times a day. I was hoping someone knows why it does this, and if there is a setting I can adjust to reduce/eliminate the event.

2. Facial Detection. I would love to be able to get this feature working, but so far I haven't had any luck. The camera is about 30' from the POI, and the sceen is about 20' x 10'. Does the subject need to be much closer to the camera for this feature to work properly? I haven't been able to get it to "recognize" a face at all during daylight or night time. Any help/suggestions here would be most appreciated.

I know that I have omitted many settings, etc. from this post, but if anyone has any experience with these issues/settings I would greatly appreciate your feedback.

Thanks,


Jeff,

Please send me an e-mail, and I will send you my phone number. I am more than happy to discuss or remote desktop, etc. I also have a camera here that you can look at and see my settings.

To use face detection, 1) Enable face detection in event menu. 2) Configure the detection area. Either use "non-detection" (masking area), and don't draw a box/box off any area to ignore, or select "detection area" and draw one or more boxes where it should look for faces.

3) Adjust the sensitivity.

I also recommend turn on the "detection result overlay" when configuring. When the camera detects a face, it will put a blue frame around the face.

4) Schedule. Make sure activation time is set to "always", or enter a specific schedule.

The trouble people run in to is getting the areas confused. The default is non-detection (mask) with no box defined - meaning the whole screen is detectable. If you do draw a box, it is now not detecting there, unless you flip the mode to detection area.

I hope that helps.

As for the iris issue, can you post or send me what the exposure menu is set to, a well as the image preset option. Also, what firmware is the camera at?

I of course recommend upgrading the camera to the latest firmware as well.

Aaron Saks

National Training Manager

Samsung Techwin America

aaron.saks@samsung.com

Aaron, thanks!

How wide of an area can you do face detect?

When faces are detected, where can they 'go'? Are they cataloged / indexed in Samsung's VMS? Can they be sent to other VMSes?

Sorry, I forgot to include that info.

Here is the best information that I have. I personally have not tested each of these specs, but I have extensively used face detection.

Maximum number of faces that can be detected at one time: 35

The camera can detect faces each with a minimum of 20x20 pixels and a maximum of 340x340 pixels

Rotation In Plane: left-right slope of ±45º

Rotation Out of Plane: left-and right rotation of± 90º, up-and-down rotation of ± 30º.

When a face is detected, it can trigger an alarm event, which our VMS (SmartViewer & Samsung Security Manager) can then display specifically as face detection. SSM can be configured to have specific event actions based upon specific event types.

Playback/search of video will then be color coded to indicate Motion Detection vs. face detection vs. analytics vs. audio detection.

Face detection, as any of our analytics or motion detection can then trigger FTP image upload, e-mail notification with snapshot, recording to SD card (configured in camera) or NVR/VMS (configured in NVR/VMS), activation of camera or NVR alarm output.

Face detection and other analytics can be configured for a schedule as to when to detect. Our SSM software also allows each indivudial event and/or camera to have unique schedules as for when their actions should take place.

The indivudial face is not cataloged, rather the video event is tagged as face detection. The blue bounding box around the face can be turned on or off in the camera. In addition, it is sent as metadata, as opposed to burned in to the video, so a VMS can turn it on/off on the client side as well. (Some VMS may not interpret this correctly and burn it in. Most do see it as an overlay or don't see it. I have seen 1 that did burn it in).

It can also be used with a hybrid analytic solution or on-camera app that works with a server or cloud solution for face recognition. The benefit of the hybrid approach is that the camera detect the face, and sends the bounding box as metadata to the recognition server, which then has less work to do, as it does not need to process the image looking for a face, rather just extracting the face and comparing it to the database.

Right now, we do not have a face recognition partner - we just launched our OpenSDK Open Platform which allows multiple apps to be loaded in to our cameras.

Please let me know if you have any further questions.

-Aaron

Rotation in plain)

What resolution are faces analyzed at? CIF, VGA, HD, etc.

You note above that the face must be at least 20 x 20. Assume an average face is 6 inches (1/2 foot), that implies a PPF requirement of 40, yes/no?

If you analyze on a full 1080p stream, that would mean a maximum FoV width of 1920 / 40 = 48 feet. However, if it's VGA, that is 16 feet wide.

Can you clarify and confirm on this point?

John,

I do not know at what resolution the FD is analyzed at. Your math makes sense, but i don't have any further info beyond that right now.

I will see if I can find out.

Also, one point to update is that different models may perform differently. I was referring to the WiseNetIII series, which ranges from 1.3 Mp to 3Mp. Some of the older models I believe used a minimum number of pixel of 40x40.

Please let me know if you learn anything else.

A common problem I have seen is that people try to detect faces in too wide of a FoV. Knowing how wide Samsung cameras can cover for facial detection would be quite helpful.

Just an update to my post. After speaking with Aaron via email and viewing his info above, it seems the "angle of the face" to the camera is barely outside the working envelope of the camera feature for my application. As you can see from the picture below, when I look towards the camera - thus giving a better view of my face - the Facial Detection works (and it works every time). The blue box is the camera's "Detecion Result Overlay" - not something I placed on the picture. Unfortunately, when I turn my head towards the car/street, it's just enough to sneak by without detection.

In sum, even thought the feature doesn't work great in my application because of the camera angle, it does work great as long as you have a decent view of a face. In fact, I used the camera settings to actually send me a SMS Text when it detected a face. How? simple. Tell the camera to "send an email" upon detection. I have AT&T for cell service, so in the camera email settings I entered the following: my10digitcellnumber@txt.att.net. This way, I get a text notification instead of an email whenever the event occurs. Other carriers are similar - just do a quick Google Search and you'll find the correct way to "email your SMS" service. I don't receive an MMS text (including the picture of the event that you would otherwise see in a normal email notification). However, I do get a text with the subject line of my choosing along with the date/time/camera IP address of the event. I do this because I set a different tone when I receive a text, and most importantly, I receive the text faster than I do normal email...typically within 15 seconds of the event. As Aaron stated above, you can set time restraints on the notifications so you only get them when you want them.

This is my way of using the available camera technology to create a "poor man's alarm system" without extra cabling and alarm bells. I'm literally within a nose of it working flawlessly - UGH!

Jeff,

Good feedback. Thanks for sharing!

Btw, it would be neat for camera manufacturers to include 'people detection' as well as face detection. This is what NICE is doing with Suspect Search (plus the hard to believe matching). In your case, 'people detection' would likely alert accurately even with you in a profile pose to the camera.

Face detection is mainly used indoor when your camera is very horizontal because you don't have a heigth angle (low ceiling). Most of the time analytic requires a minimum of 25/30° Tilt to work (typical people or car tracking engines)

With Face detect you can detect with 5 or 15° angle, so more dedicated to indoor app to double check access control. Obviously Face Detect won't work at night, with low contrast, in too noisy environment ( the famous stormy moonless night stress test) I advice to start analytic directly by the most dificult situation, to save time, because it will be always easy to get not so bad results with 5000 lux in a quite place.