🎥 Deep Learning Streamer (DL-Streamer): Build a Full Video Analytics Pipeline in One Line of Code

[biapalmeiro]

In this demo, Guy Tamir gets a fast but very deep walkthrough of how DL-Streamer enables developers to build a full, hardware-accelerated video analytics application using a single GStreamer pipeline command. Below is a summary of how the pipeline works, why it matters, and how it connects to the broader Open Edge Platform AI Suites and Libraries.

Video: https://www.youtube.com/watch?v=LGABswMnRco

It shows how to build an end-to-end application capable of:

• Ingesting video (camera feed, file, RTSP)
• Decoding / decompression
• Format conversion + image pre-processing
• Running inference (detection, classification, general inference)
• Post-processing (e.g., drawing bounding boxes)
• Rendering, encoding, or storing the output

By following only 4 steps you can get ~100 FPS real-time analytics, streaming from the internet as the result.

All of this is expressed as a single GStreamer pipeline command, composed of standard GStreamer elements plus DL-Streamer’s analytic elements.

Key Components

GStreamer (base pipeline framework)

• Handles streaming, decoding, format conversion, output

• Free, open source, widely used

GVA (GStreamer Video Analytics) elements provided by DL-Streamer

• gva-detect → object detection

• gva-classify → classification

• gva-inference → general inference

• gva-watermark → draw bounding boxes, labels, overlays

These are the building blocks for building highly customized pipelines.

Why DL-Streamer Matters

1. Hardware-accelerated performance on Intel platforms

The video explains the challenge: each pipeline stage—decode, preprocess, inference, postprocess, encode—can run on CPU, GPU, or NPU.
But mixing devices usually means heavy memory transfers and inconsistent programming models.

DL-Streamer solves this by:

• Using Intel iGPU video decode/encode accelerators

• Utilizing CPU/GPU/NPU based on optimal workloads

• Abstracting device-specific code behind GStreamer plugins

• Minimizing memory copies between pipeline stages

This gives developers high throughput (shown achieving ~100 FPS) with minimal boilerplate.

2. Unified programming model

Without DL-Streamer, developers must write different code for CPU, GPU, and NPU runtimes, plus Memory transfers and Media decode/encode.

DL-Streamer unifies all of this under a single pipeline syntax and plugin set.

3. Built-in AI model support

The video demonstrates YOLO model download + INT8 quantization automatically, making first-time setup very fast.

Main Takeaways for Edge AI Developers

• You can build production-grade video analytics pipelines using one declarative line, no device-specific code.

• Hardware acceleration is essential for multi-channel or real-time pipelines—DL-Streamer makes this accessible.

• Optimal device selection (CPU/GPU/NPU) is handled behind the scenes, but developers still retain control if needed.

• DL-Streamer complements the broader Open Edge Platform AI Suites, especially Smart City, Retail, and Manufacturing video analytics workloads.

• Perfect for developers building:

  • Smart NVRs
  • Multi-camera safety/security systems
  • Industrial inspection pipelines
  • Retail computer vision
  • Video search and summarization apps

Looking forward to your pipeline experiments, performance results, and any requests for new DL-Streamer elements or sample applications!

---

👉 Learn more about DL-Streamer
👉 Visit the Open Edge Platform Playlist for more demos