Cameras and motion

wkearney99

Senior Member
Well, dang. Figured I'd finally be able to point my cameras at a wide swath of my driveway and exclude the adjacent sidewalk and street. Worked great, no false alarms by pedestrians or passing traffic. Awesome!
Until a few nights later when someone drove by in the middle of the night. After they passed the moving reflection of their tail lights was picked up by the cameras as motion. Smack in the middle of the glass on the driver-side door of our car parked in the driveway. An area we specifically want to monitor.
https://youtu.be/P2mbT8TDmHQ
So now I've gotta figure out a whole 'nother set of angles that won't include reflected lights from vehicles passing.

I've got another one down in my workshop that trips when anyone passes along outside and breaks the light from a floodlight from a neighbor's house.  Just enough shadow passes the door to trip it.
Grrrr....
 
Yup, everyone hoped that cameras would be this cheap and wonderful AI and motion sensor, but they have problems too, as I'm sure Joshua Brown realized here:
https://www.theregister.co.uk/2017/06/20/tesla_death_crash_accident_report_ntsb/
My Ring Doorbell cam. triggers every morning and night when my landscape lights go on or off, and also when bugs are attracted to its IR light.
 
Maybe cameras combined with cheap LIDAR someday will be the answer, but cheap LIDAR doesn't exist today.
 
i have been playing around with automotive grade lidar sensors and object tracking/classification software for the past couple years. way too expensive but really neat.
 
When I tried to use video cams to detect motion outside I had false alrms caused by clouds passing in front of the sun. Nothing I tried with masking and adjusting sensitivity helped and I gave up.
 
Mike.
 
I figured as much.  Searching a bit elsewhere gives the same impression.

Clouds and shadows, yeah... but how about SNOW?  From the above referenced thread.  Oy.

I do have a few places where less-than-psychic camera/motion detection could be used, so they're not an complete waste.  

Which leads me to revisit the idea of having a continuously recording camera out there instead.  Better to have and not need than miss, I suppose. 

One... more... thing... on my to do list.
 
 
Also gives me some pause as to whether or not I'll be able to use some of these for monitoring on our boat.  I was planning on aiming them away from motion, that won't be too difficult.  But shadows and passing light, hmmm...  The inverse is true for them at night as there's not going to be passing headlights.  But armed status while away during the day could present the same hassles I have with a door at the house.
 
Well, false alarms during the day are a lot less annoying than those at night. 
 
I have been playing with home assistants tensorflow module to do object detection.   Person detection seems to work fairly well with the default models.   I really wanted  it to  detect  packages on my front porch but none of the default models detect packages or boxes, so I am hoping to train my own model.    On my core i7 it takes about 1Ghz of one core for CPU for one camera,  also playing with getting GPU setup to  see if I  can optimize things better, but some people are running this on a raspberry pi.
 
wkearney99 said:
I figured as much.  Searching a bit elsewhere gives the same impression.

Clouds and shadows, yeah... but how about SNOW?  From the above referenced thread.  Oy.

I do have a few places where less-than-psychic camera/motion detection could be used, so they're not an complete waste.  

Which leads me to revisit the idea of having a continuously recording camera out there instead.  Better to have and not need than miss, I suppose. 

One... more... thing... on my to do list.
 
 
I have the cameras in my sun lit garage set to record on motion but they don't alert me in any way. When I do review the recordings it's plain to see if it was the sun or a spider that triggered the recording.
 
I also have  Bosch motion detectors that have been flawless when it comes to not false alarming for any reason.
 
https://www.amazon.com/Bosch-BDL2-WP12G-TriTech-Motion-Detector/dp/B0056Z3W6I
 
I have had just one false motion detection in five years from the Bosch and I'm guessing that it was a bug.
 
Mike.
 
I think it is unrealistic to think you can tweek the detection software to perfectly capture the needed video without false positives. That's why they make large hard drives. Use motion detection for sure, but expect false positives. It's just taking up some drive space. You really don't even need to review it unless there is a event that raises suspecion.
 
sic0048 said:
I think it is unrealistic to think you can tweek the detection software to perfectly capture the needed video without false positives. That's why they make large hard drives. Use motion detection for sure, but expect false positives. It's just taking up some drive space. You really don't even need to review it unless there is a event that raises suspecion.
 
That's pretty much how I treat it. It uses a lot less disk space than recording 7x24 and is almost as likely to capture an intruder.
 
Here configured cameras facing the house and house out - overlapping views.  Motion detection was outdoor PIRs, under the driveway and vibration sensors.
 
Over the years the cameras went from analog to IP SD to IP HD with POE and using a 28mm lense.
 
Tinkered with camera boards mounted inside of new older outdoor domes.  Intially boards were stacked then over time went to single 38 mm square boards.
 
And a little bit of motion events from the camera views.   Still was not perfect.
 
Very quiet at night though and typically it was some coyote(s) wandering around at 3AM. 
 
There's also Google's AIY vision stuff that could be interesting.  Well, except for the 'sell your soul to advertisers' downside.

Are you running your own tensorflow rig to process the video?

 
 
No.
 
Did just find a plugin for RPi to test using tensor flow.
 
Still using ZM (> 10 years) and now testing Shinobi on Ubuntu 18.10 64 bit.
 
Interesting finding a comparison between Shinobi and ZM over here ==> features
 
I am not optimistic of the outdoor stuff using video algorithms. 
 
It is getting better as the last camera boards I played with are little computers with faster CPUs, more rom and ram...
 
pete_c said:
No.
 
Did just find a plugin for RPi to test using tensor flow.
 
This link below looks like a good compromise solution to deal with the CPU issues with Tensorflow.   He sets the scan interval out to a week and just calls the scan manually when blue iris detects motion.
https://community.home-assistant.io/t/improving-blue-iris-with-home-assistant-tensorflow-and-pushover
 
The recommended Coco RNN model on the HASS website seems to be the best one in my testing for CPU usage.   It will detect people pretty well and supports cars and some other stuff (and apparently umbrellas which is what mine falses on  all the time :) ).  
 
Thank-you Wuench. 
 
Been tinkering with HA but have not tried video stuff on it. 
 
Still running ZM here and now testing Shinobi.  I can probably do the same using the Shinobi feeds.
 
Back
Top