AMD Video Analytics Framework

日本語

A GStreamer-based video analytics framework for AMD APUs. Decoding of camera feeds and video files, inference such as object detection, drawing of detection results, and encoding all run on the GPU; only metadata such as detection results returns to the CPU. Targets Strix Point and Strix Halo; see the current validation matrix for tested configurations.

Demo

For details and launch instructions of the NVR demo, which records multiple cameras while displaying detection results, see the NVR demo.

Features

• Processing that stays on the GPU: Pixels are never copied to the CPU between decoding, preprocessing, inference, drawing and encoding.
• Built from GStreamer elements: Use object detection, classification, instance/semantic segmentation and pose estimation in gst-launch-1.0 pipelines.
• Bundled DL Streamer elements: Tracking (gvatrack), multi-stream batching and metadata conversion are available under the same element names as in DL Streamer.
• Result output: Publish detection results to JSON Lines, MQTT or Kafka. Annotated video is hardware-encoded and can be saved to a file or served over RTSP.
• C++ SDK: Embed RTSP camera ingest, recording, and recording search and playback into your application without depending on GUI libraries.
• Samples: A 16-camera NVR demo, a GPU-rendered resource dashboard, and a tool that serves video files as RTSP cameras.

Requirements

Ubuntu 24.04 or 26.04, Strix Point/Halo, host ROCm 10.1.0 and MIGraphX 2.18.0. See installation and the validation matrix; not every target combination has completed hardware acceptance.

Documentation

The documentation is available at https://anjn.github.io/amd-video-analytics-framework/.

• First-time users: Installation → Quick start
• Building pipelines: Building pipelines
• Embedding in applications: Embedding the SDK in your application
• Design, all properties, validation results: Developer documentation

License

The code in this repository is under the MIT License. Bundled third-party sources (parts of DL Streamer, sqlite-vec, etc.) are subject to their respective licenses; see THIRD_PARTY_LICENSES.md for the list. Model weights and sample images are not included. If you use them, check their respective terms of use.

Reproducible host-SDK setup

The current prerequisites are Ubuntu 24.04/26.04, ROCm 10.1.0 and MIGraphX 2.18.0. The former bundled ROCm 7.2 workflow is retired. See installation and validation status. Models are generated from a fresh checkout; users supply the 1440p videos.