Using deepstream while merging multiple rtsp streams

Hi, I’m currently trying to find out whether I can convert my opencv based program into deepstream. The pipeline I am using is currently using ROS, and on a high level:

  • get frame from each camera using gstreamer/opencv, publish to topic
  • a processing node applies perspective transform and stitches the images together
  • run detection model
  • display stitched image with ROI, additional UI elements, and bounding boxes.

I understand i can run detection on seperate rtsp streams, and seperately draw bounding boxes in each one, and even apply perspective transform using VPI VPI - Vision Programming Interface: Perspective Warp.

But is there any way I can for example apply perspective warp to each of the streams then “combine” it into one stream?

What kind of camera? CSI camera? USB camera? IP camera with RTSP streaming? Or others?

Have you implement the algorithm with CUDA?

Please check whether Gst-nvdewarper — DeepStream documentation 6.4 documentation meets your requirement?

Which Nvidia platform will you use?

  • I will be using an IP camera via RTSP streaming.
  • I have not implemented the algorithm with CUDA, only in c++ using OpenCV
  • I will check out Gst-nvdewarper, thank you for the info.

I am also going to be using an Jetson Orin agx

after taking a look at gst-nvdewarper and its github page, im not sure if its exactly what Im looking for. It appears to go from one source (e.g. fisheye image) and splits it into multiple through dewarping. I’m going in the opposite direction, where I have 3 cameras with some overlapping area, and I essentially want to create a panoramic view of through stitching the 3 images/streams together

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

There is no stitching algorithm implemented in DeepStream SDK. You may need to customize the algorithm and insert the algoithm into one GStreamer plugin. The CUDA accelerated algorithm is preferred to get better performance while the openCV is acceptable too.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.