human detection on the mower

[xjust]

Hi there,

I just tested deepstack+aitool+blueiris for security purposes on all the cameras installed on my DVR.
The results are amazing for human detection. After little optimization there are 0 false alarms.

Im thinking this can be used in ardumower for obstacle detection using a camera mounted in front of the mower.
Takes about 50ms to analize photos using gpu acceleration.

 

[bernard]

NEED SOME KNOWLEDGE TO TEST.
But this feature is actually include into Piardu510 ,Only tested for less than 2 hours but without particular issue except the PI Power CPU usage (More than 50 % and no html video streaming on Home assistant) on a raspberry pi 4 with Picamera.
Piardu510 use an AI Tensorflow to detect PERSON,DOG etc and automatically change mower heading according the location of person in the frame to avoid it

 

[xjust]

If there is wifi coverage might be better to offload the AI processing externally in this way there is no more cpu load on the pi4
also a good ideea is to mask some of the left/right side of the image so there is no change in direction if the mower is not heading straight to the person/dog

 

[bernard]

If there is wifi coverage might be better to offload the AI processing externally

It's a mower and wifi is never perfect everywhere in a garden ,so i am not really sure !!
Using PI onboard computation you can run without WIFI and if WIFI is present VNC VIEWER can clone PI Screen and the image have no latency.

Mask left and right is really not a problem .
The main problem is that it's not possible to know the person distance (It's simple video and not depht one)
But like i say it's only DEV

Actual setting are :
#Setting for Vision
personDetectNewHeading = 30
personDetectNewHeadingDuration = 2
personDetectMinScore = 78

Score detection , heading relative and the duration mower drive before return to his lane.

More on next SUMMER.

 

[xjust]

its difficult if you have kids in the garden.. otherwise you can relate the distance to the person by the height of the detected object
cheers