AMD Video Analytics Framework¶
A GStreamer framework for analyzing camera feeds and video files on AMD APUs (validated on Ryzen AI 9 HX 370 / Radeon 890M). Decoding, inference, drawing and encoding all run on the GPU; only metadata such as detection results returns to the CPU. By linking GStreamer elements with gst-launch-1.0, you can run object detection, tracking, classification, segmentation and pose estimation, and output the results as video files, RTSP streams or JSON.
What you can do¶
| Use case | What to use |
|---|---|
| Draw detection results on video or camera feeds, then save or stream them | GStreamer elements (amddetect, amdwatermark, amdencode, etc.) |
| Send detection results to other systems as JSON or over MQTT/Kafka | gvatrack, amdmetapublish |
| Build analysis, recording and playback into your own application | C++ SDK (CMake package) |
| Try multi-camera recording, monitoring and person search | NVR demo |
Reading order¶
- Installation: Build the dependencies and verify that the GPU works.
- Quick start: Prepare a model, draw detection results on a video and save it.
- Concepts: Where the data flows, and the roles of elements and metadata.
- Depending on your goal, continue with Building pipelines, Embedding the SDK in your application or the NVR demo. If you used DL Streamer or DeepStream, read Migrating from other frameworks first.
Internal design, validation results and the full property list are in the developer documentation.