Repolode
投稿する
sherlockchou86

sherlockchou86/VideoPipe

A C++ video analysis framework that chains stream reading, model inference, tracking and output as pluggable pipeline nodes.

代替できる製品:NVIDIA DeepStream

GitHub で見る公式サイト
スター
2,984
フォーク
470
未解決の Issue
4
最終コミット
8 か月前

これは何？

VideoPipe is a C++ framework for video analysis and structuring. It is built around a pipeline of independent nodes that can be combined in different ways, so developers can assemble applications such as face recognition, image search, object tracking and traffic or security behavior analysis without writing the plumbing themselves.

The framework handles stream input over protocols like RTSP, RTMP, UDP and files, decoding and encoding through OpenCV or GStreamer with hardware acceleration, and on-screen display of results. Inference uses OpenCV::DNN by default and can be switched to TensorRT, Paddle Inference, ONNX Runtime or a custom backend. Recent versions also integrate multimodal LLMs served through Ollama, vLLM or OpenAI-compatible APIs.

VideoPipe supports multi-level inference, IOU and SORT tracking, line-crossing and parking detection, and pushing structured data to Kafka or sockets. It builds on CMake with optional CUDA, TensorRT, Paddle, Kafka and LLM modules, and runs on CPU when those are disabled.

主な機能

  • 01Reads streams over RTSP, RTMP, UDP, files and applications.
  • 02Supports pluggable inference backends including OpenCV::DNN, TensorRT, Paddle Inference and ONNX Runtime.
  • 03Performs multi-level inference for detection, classification and feature extraction.
  • 04Tracks objects with IOU and SORT and analyzes line-crossing and parking behavior.
  • 05Integrates multimodal LLMs via Ollama, vLLM or OpenAI-compatible APIs.
  • 06Pushes structured data to Kafka or sockets and overlays results on frames.

ユースケース

  • Building traffic incident or violation detection from camera feeds.
  • Prototyping face recognition or vehicle tracking on edge devices such as Jetson.
  • Running video structuring pipelines on CPU-only machines for testing.
  • Sending detected events to Kafka for downstream processing.

トピック

#computer-vision#video-analysis#deep-learning#object-detection#object-tracking#gstreamer#opencv#multimodal-llm#ai#behaviour-analysis

よくある質問

VideoPipe とは？
A C++ video analysis framework that chains stream reading, model inference, tracking and output as pluggable pipeline nodes. VideoPipe is a C++ framework for video analysis and structuring. It is built around a pipeline of independent nodes that can be combined in different ways, so developers can assemble applications such as face recognition, image search, object tracking and traffic or security behavior analysis without writing the plumbing themselves.
VideoPipe は無料で使えますか？ライセンスは？
VideoPipe は Apache-2.0 ライセンスのオープンソースです。商用製品で使う前にライセンス条件を確認してください。
VideoPipe は今もメンテナンスされていますか？
半年以上コミットなし。VideoPipe の最終コミットは 2026年2月25日 です。
VideoPipe は何に使えますか？
Building traffic incident or violation detection from camera feeds. Prototyping face recognition or vehicle tracking on edge devices such as Jetson. Running video structuring pipelines on CPU-only machines for testing. Sending detected events to Kafka for downstream processing.
VideoPipe はどこで入手できますか？
ソースコードとインストール手順は GitHub にあります：https://github.com/sherlockchou86/VideoPipe

概要はプロジェクトの README をもとに AI が生成しています。GitHub の数値は毎日更新されます。

関連プロジェクト