Group Actions
Engagement
  • Author Author: sakthi.1260
  • Date Created: Date Created
  • Views 708 views
  • Likes 7 likes
  • Comments 4 comments
Related
Recommended

Bluetooth Unleashed : AAPSAD #3 : The Fall

sakthi.1260

Hi guys,

 

There's been a lot of confusions, lots of algorithms, lot of choices for the face analysis part. Spent a lot of time over it, learnt a couple of new things, so what else it was time to test things out.

 

I started with the traditional way, template based matching had to take around 35 pics and my beard for better accuracy, after a bit of enhancements and the discovery was, I'm really bad with Expressions image it wasn't working the way I want it to.

image

So the next option was to make a standalone AI to do the job, so I called a friend with knowledge in Machine Learning and OpenCV, he had already had an python script for the same, we tried and it was working good.

image

But the problem was, the script was designed to run on a high spec computers and our poor hero had to give in all his abilities to run. There was like 5s delay to move into the next frame. The ACT LED stood still and had to pull the chord to stop the process.

image

 

With that sort of system load, video playback (the visuals and audio part) would never work.

 

There has to be an way, I'm looking into cloud services for the same, started working not sure which one will work out

 

Will the hero rise back??

To be continued...image

 

Cheers,

  • ,

     

    Each pixel used has a cost.  Try to limit the area that you are processing and the resolution that you are using.  I have done a lot of image processing and have found that highly optimized algorithms always beat general purpose functions.  Also, be sure to utilize integer math where ever possible as this is the most optimized calculation.

     

    Good luck as you regroup and figure a way to achieve your goals.

     

    Gene

  • If you do a little research, you will find that the higher quality recognition algorithms use a set of control points for facial features that they use as cues for specific purposes.

    Take a look at those control features, especially the eyes, nose and lips. You should find that you can improve your reliability and ability to resolve more facial characteristics.

     

    DAB

  • Hi!

    Good to see the progress, nice project : )

    If you're only applying processing per frame using OpenCV (i.e. only looking for features in still images), then there are a few tricks possible. One idea is to reduce the image resolution (since you likely don't need HD), and another is to choose not to process all frames, e.g. just process one frame a second, or get the Pi camera to take a snapshot once per second only. It should speed things up significantly.

  • Nice update! Perhaps limit the frame rate or only trigger the script based on some other input like a button press or motion sensor?