yolov7-daemon, secondary object detection to eliminate false event notification

[feelingwalnut]

never could get secondary detection to work within motion so this should run on_event_end.
this came to fruition after digging through a conversation here on github about similar issues with secondary_detection and a desire to keep everything in memory to save R/W on the disk unless it is "worth" keeping.
all made in chatgpt.
https://github.com/feelingwalnut/yolov7_daemon
daemonized the detection, this version gets a directory passed to it when the event ends, searches for the image to process, makes a decision to dump it or move it and, if set up, will message the user with the annotated version of the image.

or i've also a wrapper script for on_picture_save. this processes the image and decides to dump/move and messages similarly.
https://github.com/feelingwalnut/yolov7_detect_wrapper

i'm not providing detailed setup instructions, just saving some time/effort if you want a similar solution.

the wrapper has an issue where if the event is long, it may process the image and take action while files are being saved, leaving crumbs.
the daemon waits until event end and has a queue.

[xj25vm]

I have implemented something similar with Yolov8, except it also decides if the resulting movie file should be kept or not. One issue your script might have is that if the event is long - for example there might be wind constantly moving vegetation, or a person moving about in front of the camera for long periods of time - if the processing is only triggered at the end of the event, users might get alerts a long time after the relevant motion occurred. This might be ok depending on your application though

[feelingwalnut]

I was looking to eventually move to newer models but am not very good at this stuff so couldn't figure it out.

This system does as you describe, deciding to purge from memory or move to storage depending on criteria like category and threshold.
The issue you mention is accurate, if it is a long event, it won't notify until end, I did this because it was cleanest solution to avoid move/delete of long run events before finishing.
The delay is fine for my use but it could be modified to process image when motion first detected and then move/delete at event end. I might do that in the future.

My system, a j5040 Pentium, doesn't process images quickly so there is a minute or two delay regardless.

[xj25vm]

I mentioned about long running events because I just had an issue with one of my systems in the last few days. The weather has been windy, and I discovered that events kept on running for hours because of the constant movement from wind - so notifications can be delayed a long time. If you can arrange to have the cameras covering an area without vegetation, or are able to mask off the vegetation completely in Motion, then your script could still work with satisfactory results though.

[feelingwalnut]

Oh, I've experienced the same. The tough balance of motion sensitivity, masking areas, private masking objects like a parking spot so the detection doesn't trip all night over my parked vehicle and such through motion to tailor triggering.

I also use video passthru to hide the private mask from video file and to provide audio