I would rather have a clear picture of what is going on vs a picture of a persons arm.... The performance is very hardware dependent I analyze up to 900 images per trigger with a high accuracy 90%+ applied to get the clearest image in the alert. My servers only take 1-2 seconds to fire off the alert after the initial alert is sent to CPAI. To me that is fast enough as I cannot dial the police or get to my door that fast. But the alert time is not so important on my config vs positive identification of the subject material. I also record all streams 24/7 in case something gets missed.
In the case of the op 250ms per image at 10 images would only be 2.5 seconds for positive identification. It takes longer for the alert to be sent than the delay in analyzing it depending on the network/internet speed. The main thing I was going off of was the op's want to know why the "person" detected ended up in the canceled alerts, adding giraffe to the To Cancel line would help alleviate that, giving the AI more time to finish analyzing the images given to promote an active alert vs a canceled one.
My suggestion to the op would be to increase the amount of images analyzed slightly, reduce the analyze time to fire at 250ms, along with increase the accuracy quite a ways. If he/she is still getting canceled alerts I would add giraffe to the to cancel line.
The bottom line is there is always give and take on a case by case basis, Because every use case is situational. All we can do is make suggestions to help them along
