jakaskerl Thanks a lot.

I don't have any information about training another model and use it for that. I don't know if Luxonis provided that.

Cam could be located on the ceiling and at the very extremity of the space, in order to be oriented (looking at) the whole scene. Wouldn't be from top view but a bit like 3/4 oriented toward the center of the room. I think it could detect silhouettes…

Actually, this is exactly as in the youtube video example I posted in my first message here.
Btw, @erik , an we find this specific example somewhere ?

Actually, I wouldn't have more than… let's say…. 25 points. Light.

  • erik replied to this.

      erik hi @erik and thanks again for your support.

      Actually my worry is more about how to use this kind of view angle with the best possible algorithm for detecting + calculating distance between people.

      The example I mentioned and which is included in your examples set for counting people with a cam on the ground works very fine. I tweaked it for sending messages over osc over udp to another program I coded, I send only distances lower than a specific value. I can easily track new "encountering" only (people coming nearer than a radius limit) worlds very fine.

      Now I need to scale it up a bit.

      Ceiling is 5m

      I'd really like to merge the example up above with the other one.

      • erik replied to this.

          Hi julienbayle , Sorry I am a bit confused - which two examples are you referring to (that you'd like to merge)?

            I'm sorry about the confusion.

            Basically, I already used gen2-social-distancing example. It was working.

            I'd need to use it for measuring distance between people but with a camera located at the ceiling, oriented like in the youtube video I posted.

            This is just what I'm trying to do.

            I'm still not sure about the cam I'd need for covering the widest area. I need to order it asap as I'm in Europe.

            French distributor didn't answer yet.

            erik I hope you could help me to know which cam I need to target for this. French distributors can get easily oak1 and oak d pro.

            Would these models could work for this area width and considering low light condition?

            I'm lost (in the middle)

            • erik replied to this.

                Hi julienbayle ,
                For the social distancing demo, you'd need depth. So you should look for some OAK-D * product that's Wide FOV. Then you can select USB or POE connectivitiy, and whether it's Pro (night vision/active stereo). For EU distys, you can found them here: https://www.luxonis.com/distributors
                Thoughts?
                Thanks, Erik

                    erik

                    Thanks a lot.

                    So D required. ok. W is not required ? I think it is.

                    USB here ok. and is Pro required for low light ? does the example we mentioned can work with night vision as it is ?
                    Is the laser/stuff feature to get more accuracy in low light environement is "visible" ? Does it fire laser everywhere in visible spectrum ??? (laughing at me writing this… looks like star wars battelfield)

                    If I understood correctly, if Pro brings a better sense of people tracking in low light conditions, I think I need:

                    OAK D PRO W

                    Not sure about IMX378 VS OV9782. The second one brings the same sensor for the color camera than the one for the stereo pair). but 12M vs 1M. Which one would you suggest for best analysis (same trained model) and WIDE picture ?

                    I think with these answers… I'm okay for buying it now

                    • erik replied to this.

                        Hi julienbayle ,
                        Yes, "W" (Wide FOV) is required if you want the widest FOV (to cover the most area possible). Illumination LED (that comes on OAK) can light up the area to eg. 3m. Anything above that it won't be able to illuminate much. It "fires" lasers at 940nm, so it's not visible to people🙂
                        You likely won't use object detection models that are larger than 1MP (as that would be very slow on RVC2), so 12MP would only make sense if you want to do additional things, like age/gender estimation, face recognition, etc. Thoughts?

                            erik This is very clear.

                            For that application, specifically, I don't need more than 1MP for this.

                            So, OAK D PRO W with OV seems to be my best bet. Right ?

                            gosh, thanks so much.

                            • erik replied to this.