Person Tracking AWS Rekognition Video (Flickering Bounding Boxes)

Question

Hi everyone, I am working with AWS Rekognition for video analysis of person tracking. I have successfully gathered all the bounding box data as well as the facial landmarks data per frame. So what I have done is grab the frame at a given timestamp and apply the bounding boxes and landmarks to the frame. However, when stitching the video back together I noticed the bounding boxes and facial landmarks flicker between frames. I am pretty sure this is due to the fact that not every frame in the video has data populated for bounding boxes and facial landmarks. Is there any solution to make the bounding boxes and landmarks not flicker when stitching the video back together. Any assistance is greatly appreciated.   
  
Thanks in advance. I have attached a video of the current issues I am encountering for explanation's sake.  
  
Edited by: brussell152 on Nov 13, 2019 10:29 AM  
  
Edited by: brussell152 on Nov 13, 2019 10:39 AM

Answer

Hi,  
  
Yes, detections are generally not output for every frame, so if you are rendering boxes on top of frames, you will need to add some persistence logic. The simplest approach is to render all the boxes for a given output timestamp at once, and leave them drawn until the next timestamp, and then do a full redraw. Another approach is to interpolate (tween) between boxes for upcomming timestamps. You can use the person ID field for this.  
  
The Rekognition video console demonstrates examples of these approaches.  
  
Hope this helps,  
Jonathan

Person Tracking AWS Rekognition Video (Flickering Bounding Boxes)

Relevant content