I'm starting to look at using a second AI that only runs on current detections in an attempt to reject false positives
Reading "It's Not Perfect" I think this sounds as an interesting idea. I also think it could be worth trying to change the detection model/type for the second analyze to something else not based on ssd. Even if it is heavy loading your cpu, you could try to run YOLOv3 for the second run, it might be that would help (and then it is just a matter of finding enough powerful cpu:s )
I will try to run your images with those false alarms through my YOLO analyzer and see what happens
After testing your bogus images, the result is:
The image with the bird is the only one where "person" was detected when running it through my RPi3 with Tensorflow JS Coco SSD node in Node-RED. For the other images, no false detection at all
None of the bogus images resulted in false detections of "persons" when running them through:
NVIDIA Jetson Nano: jetson inference engine with ssd-inception-v2 using the GPU of the device
Lenovo laptop w debian: YOLOv3 handled by Python
So for a second run, I think using either would filter out those false detections. I think I would go for and try out YOLOv3
EDIT: Forgot to mention, the analyze of your bogus images took around 0.17 sec in the Jetson and around 0.78 sec in my Lenovo laptop
Thanks for doing this and the info! I've got bogged down in other things,since my current false alarm rate is low enough to not be a bother, but this motivates me.
In my node-red I filter detections to reject people walking by on the sidewalk and only alert when they get a certain distance towards the house. Sending the image to a Nano for final verification before the alert and letting the Nano do the alert would be a minor change once I get the Nano AI running.
Hello krambriw,
Thanks for you sharing about your node , but now I come across one question,please help me ,thanks in advance.
I can input your flow and it can deploy correct,but it can not display with broken picture icon.Like above GChapo mentioned.I use AXIS IP camera , do you know what's the reason?Thx.
Well, first thing you have to do is to find the url you need for the axis camera. This url must work in a browser so this is what you have to find out as first step. Maybe requires user and password? All this you need to figure out first and just using a browser like Chrome to test. Anyway, be happy you have a high quality axis camera
@Krambriw Thanks for your quick response.I will find url use below Steve-Mcl sharing link. I want to know if I find the correct url ,next step is I need install Motion as you mentioned before and config motion.conf .Do I need run Motion background after linux system boot finish?
@krambriw In fact.I used ffmpeg run success as BartButenaers suggest method in " [How to display CCTV camera in dashboard (RTSP)]" ,But I found CPU consumption is high(you mentioned to him in that discussion), and the video picture is stuck sometimes ļ¼ So I want to try your suggested method.
Hi,
Like Walter mentions, you should have a look at that discussion about the nodes from @kevinGodell. Kevin is a video magician, while my rtsp flows were only a quick and dirty experiment. At the time being we had no decent video solution in Node-RED, so I tried to have at least a minimal setup. But Kevin's stuff is MUCH better!!
Bart
)
