Close

Best practices to optimize for people detection latency?

Learn how to turn an Android phone into an IP camera security system with Sighthound, download automation scripts & more.

Moderator: Staff

no avatar
luma
 
Posts: 14
Joined: Sat Nov 28, 2015 11:54 am

Best practices to optimize for people detection latency?

by luma » Sun Nov 29, 2015 9:39 pm

I am using Sighthound in conjunction with a handful of home automation technologies with the hopes of utilizing the pedestrian detection algorithms as a motion sensor that won't be triggered by my cats and dogs.

I have very little concern about image quality, storage, or any of the traditional security camera needs. What I want is to pull a stream from a camera, determine if a person is in a specified area, trigger an external program, and do that as quickly and reliably as possible.

In testing my system I am seeing roughly 10 seconds between an action taking place and Sighthound triggering an action ("Run command"). For some use cases this is OK, but it'd be nice to get this done faster for things like turning on room lights when a person enters as 10 seconds is a long time to be standing in the dark (and quite a bit longer than it'd take to walk over to where the switch is).

Given that requirement, are there any best practices for setting up Sighthound for low latency person detection? For example, I have two 1080p cameras. Should I be pulling from low resolution feeds to make the processing happen more quickly, or does the resolution not impact the processing speed by much? Would low resolution negatively impact accuracy (false positives or negatives)? Is there a performance difference between MJPEG feeds and RTSP sources?

Are any particular system resources going to play a significant role here (CPU, disk I/O, etc)? In my case the CPU utilization isn't being maxed out on the host system, memory utilization is reasonably low, and disk I/O doesn't appear to be congested.

I'm curious to hear if anyone at Sighthound or some knowledgeable users have dug into people recognition performance optimization and what, if anything, has been learned.

User avatar
Devin
 
Posts: 719
Joined: Wed Mar 12, 2014 5:48 pm

Re: Best practices to optimize for people detection latency?

by Devin » Wed Dec 02, 2015 1:10 pm

Hi luma,

A small amount of delay is normal, though the types of rules you have set, as well as the events themselves, can create a shorter or longer delay. This support article describes how different rules and events can affect the delay you see in actions: https://www.sighthound.com/support/ther ... eing-taken. Resolution and other image settings shouldn't have any significant impact on the delay, unless your computer is severely overburdened in which case you would also be experiencing other issues.

Regards,

-Devin

no avatar
luma
 
Posts: 14
Joined: Sat Nov 28, 2015 11:54 am

Re: Best practices to optimize for people detection latency?

by luma » Thu Dec 03, 2015 8:33 am

That is exactly the sort of detail I was looking for. Thanks Devin!


Return to Hacker's Corner
cron