Hi,
We have deprecated omx plugins so please use nvv4l2decoder.
Since OpenCV processes frame data in BGR format, and BGR is not supported by hardware compoments in Jetson chip, so we would need to use CPU to copy BGRx from NVMM buffer to CPU buffer, and then convert to BGR. This would take significant CPU sage. Your pipeline is optimal to get frame data in BGR.
For further enhancement you may check if you can use GpuMat and apply CUDA filters. Here is a patch for reference:
LibArgus EGLStream to nvivafilter - #14 by DaneLLL
We also have VPI functions:
VPI - Vision Programming Interface: Main Page
You can check if the required OpenCV functions are supported in VPI. If yes, may consider use VPI instead of OpenCV.