Skip to content

NVIDIA DeepStream

DeepStream is an NVIDIA-accelerated GStreamer graph. Its main advantage is keeping decode, batching, inference, tracking, and composition inside NVIDIA's native pipeline and device-memory model.

Use easy-rtsp as a source boundary, not as a frame-by-frame middleman:

camera / file / custom Python source
                 ↓
             easy-rtsp
                 ↓  RTSP over TCP
            DeepStream
                 ↓
       inference / tracking / metadata

Publish from easy-rtsp

On the source host:

easy-rtsp relay rtsp://camera/live --serve edge --lan

This is useful when you want easy-rtsp to normalize camera access, reconnect upstream, or expose a predictable URL to a separate Jetson/dGPU service.

Configure deepstream-app

Add an RTSP source group to a DeepStream application config:

[source0]
enable=1
type=4
uri=rtsp://192.168.1.42:8554/edge
select-rtp-protocol=4
latency=100
rtsp-reconnect-interval-sec=10
rtsp-reconnect-attempts=-1

[streammux]
live-source=1
batch-size=1

The important choices are:

  • type=4 selects a DeepStream RTSP source.
  • select-rtp-protocol=4 forces TCP, matching easy-rtsp's local MediaMTX transport.
  • live-source=1 tells the stream muxer that timestamps come from a live source.
  • batch-size should normally match the number of enabled sources.

Run your normal DeepStream configuration:

deepstream-app -c edge-app.txt

The exact inference, tracker, OSD, and sink groups remain model- and deployment-specific. NVIDIA's reference application configuration is the source of truth for the installed DeepStream release.

Multiple streams

Publish each easy-rtsp stream under a distinct path:

rtsp://edge-host:8554/camera/front
rtsp://edge-host:8554/camera/loading-dock
rtsp://edge-host:8554/camera/parking

Then create one [sourceN] group per URL and set the streammux batch size to the source count.

One easy-rtsp process currently owns one generated MediaMTX path

Multiple independently started local servers cannot all bind the default port. For a multi-camera edge box, run one shared MediaMTX instance and have each easy-rtsp process publish to a full RTSP URL on that server.

Example:

easy-rtsp relay rtsp://camera-a/live --serve rtsp://127.0.0.1:8554/camera/a
easy-rtsp relay rtsp://camera-b/live --serve rtsp://127.0.0.1:8554/camera/b

When not to use easy-rtsp

If DeepStream can directly open the camera and all processing stays inside its graph, connect it directly. Passing frames through easy-rtsp's NumPy map() path would move them to CPU memory and re-encode them, defeating much of DeepStream's zero-copy advantage.

Publishing DeepStream output

DeepStream can create its own encoded RTSP sink (type=4). Let DeepStream own that output when the analytics graph is already authoritative. Use MediaMTX separately when you need a shared routing layer, authentication, or browser-facing protocols.

Platform boundary

DeepStream is intentionally not an easy-rtsp dependency. It is tied to supported NVIDIA GPUs, Jetson/dGPU platforms, CUDA, TensorRT, GStreamer, drivers, and specific Linux releases. Consult the current DeepStream installation matrix before choosing an image or host OS.