Skip to content

README.md

Latest commit

 

History

History
108 lines (85 loc) · 4.73 KB

README.md

File metadata and controls

108 lines (85 loc) · 4.73 KB

Project Design: VLX VisionBridge

Project Overview

VLX VisionBridge is a headless, high-performance Linux service written in Go. It aggregates multiple finite SRT/WebRTC/Media streams into a single composite live stream, broadcasted simultaneously to multiple CDNs (YouTube, Twitch, VK).

The service is designed for professional 24/7 broadcasting environments where configuration must be dynamic and resource efficiency is paramount. We are basically building a sort of obs-studio for remote VMs.

Requirements Note

  • Hardware: Multi-core CPU for FFmpeg processing, adequate RAM for media buffering.
  • Software: Modern Linux distribution (e.g., Ubuntu 20.04/22.04), FFmpeg installed and accessible. Xvfb and Chromium (optional, if using overlay HTML sources).
  • Network: High-bandwidth, low-latency network connection to handle multiple SRT/WebRTC streams and simultaneous broadcasting to multiple CDNs.

Core Principles

  • Headless First: Managed entirely via configuration files or DB entries.
  • Dynamic Reconfiguration: Hot-reloading of layouts and sources without dropping the output stream (where technically possible).
  • Resource Optimization: Sources marked as "OFF" are completely excluded from the processing pipeline.
  • Multi-Destination: Single encoding pass with multiple output clones.

Best Practices

  • Top-Most Overlay: The chromium_source overlay should be kept as the top-most layer in the filter chain to avoid performance issues associated with complex layering. It is automatically rendered on top of all other sources.
  • Layer Conventions: To ensure a stable and performance-oriented configuration, ffmpeg_source is limited to a maximum of 3 layers (IDs 0, 1, 2) which share identical capabilities. The recommended convention is:
    • Layer 1: Primary Input (e.g., GoPro).
    • Layer 0: Fallback/Placeholder (e.g., VODs).
    • Layer 2: Secondary Input / Guest.

Layer Configuration Examples

VisionBridge operates alongside MediaMTX and ChatBridge on the same localhost. It handles low-latency video ingestion and real-time scene switching via ZeroMQ without restarting FFmpeg.

IPC Audio Input

Accepts raw PCM data directly via a Unix Domain Socket. If InputPath is left empty, it defaults to /tmp/vlx_audio.sock, but it can also be explicitly defined.

{
  "id": 1,
  "active": true,
  "input_type": "ipc_audio",
  "input_path": "",
  "media": "Audio",
  "volume": 1.0
}

Folder Playlist Input

When the input path is a directory and the layer configuration includes folder_options with is_folder: true, VisionBridge treats it as a playlist. It plays all valid video files (e.g., MP4, WebM) found in the directory. You can shuffle the playlist, loop it, and insert a delay (customizable delay spacer (color or image-based)) between videos.

id: 3
active: true
input_type: "local"
input_path: "/opt/VLX_VisionBridge/data/layer3"
media: "Video+Audio"
size: 1920
x: 0
y: 0
volume: 100
folder_options:
  is_folder: true
  shuffle: true
  loop: true
  delay_sec: 5
  spacer_width: 1920
  spacer_height: 1080
  spacer_fps: 30
  spacer_sample_rate: 48000
  spacer_color: "black"
  spacer_image: ""

Network Input

Optimized for MediaMTX via RTSP/SRT. The network input_type automatically injects -fflags nobuffer -flags low_delay for true zero-latency ingestion from MediaMTX (SRT/RTSP).

{
  "id": 2,
  "active": true,
  "input_type": "network",
  "input_path": "rtsp://localhost:8554/stream",
  "media": "Video+Audio",
  "size": 1920,
  "x": 0,
  "y": 0,
  "volume": 1.0
}

Layer Control Rules

  • Rule 1: ZMQ commands MUST ONLY target ffmpeg_source layers (hardware cameras, local videos).
  • Rule 2: chromium_source layers (Overlays, Alerts, Maps) MUST be kept active: true constantly. Show/Hide logic for web layers must be handled via WebSockets/JavaScript, NOT via ZMQ, to avoid FFmpeg restarts and stream drops.

Dynamic Regia

Scenes are toggled on/off screen by moving them to x -9999 and setting volume 0.0 via ZMQ. This mechanism guarantees zero frame drops when switching scenes since FFmpeg doesn't have to restart or reload inputs.

Configuration Concepts

  • Canvas Size vs. Output Size: The fundamental drawing area for layers and overlays is controlled by input.resolution (InputSettings). The final resolution of the stream that is encoded and pushed to your destinations is controlled by output.resolution (OutputSettings).

Technology Stack

  • Language: Go (Golang)
  • Processing Engine: FFmpeg (via os/exec)
  • Database: SQLite (State persistence, Logs, Metadata)
  • Messaging: ZMQ is a mandatory dependency for real-time filter communication with FFmpeg.

See Architecture for High-Level Design details.