Queue in videostream for machine learning

The objective is to Pan, Tilt and Zoom in a video stream with focus of where a specific object (ball) is.

How can I design a queue in the videostream where the machine learning operates on whatever goes i the queue and the video is processed (1-3) seconds in the queue, whith input from the machine learning.

This would give the machine learning more time (1-3) to get a proper object detection, as objects might be blocked by other objects for shorter periods (< 1s).

Any advice, demos, papers etc. would be helpful.

Not sure but it looks to be a kind of IP camera:

Maybe you can try DeepStream SDK. If it is an IP camera, it should work by setting to RTSP source.

Hi DaneLLL,

I was a bit unclear, the system will cut out a 1920x1080 stream out of the 8000x1600 video stream, to replicate a PTZ function. But the real question is how to operate on a queue, where the streaming will operate 1-3 s after the machine learning parts.


We don’t have experience of using the camera. Are you able to see video preview by running

gst-launch-1.0 uridecodebin uri='rtsp://_RTSP_URI_' ! nvoverlaysink