GenAI · Vision-Language Models · Computer Vision

Industrial Safety Vision AI

Real-time video safety monitoring powered by a vision-language model, paired with a FLUX-based generator that synthesizes rare hazard scenarios for training.

Role
AI Intern
Org
Detect Technologies
Timeline
Jan — May 2025
Mode
Real-time, multi-threaded
Qwen VLMFLUXOpenCVStreamlitHugging FacePythonMultithreading

Motivation

Industrial sites need continuous safety monitoring, but hazardous events (missing PPE, unsafe proximity) are rare in real footage — which starves any model of training data. The project had two halves: a real-time analysis pipeline to watch live video, and a generative pipeline to synthesize realistic edge-case imagery so the system could learn what danger looks like.

Architecture

A Vision-Language Model interprets video frames against safety prompts in real time, while a separate FLUX-based generator produces synthetic training data for rare scenarios. Everything is observable through a Streamlit interface and a structured logging framework.

📹Live VideoOpenCV capture
🧵Multi-thread Framesparallel decode
👁️Qwen VLMscene reasoning
🚨Safety Verdict+ logging

Real-time path. A parallel FLUX generator feeds synthetic edge cases back into training.

What I built

  • End-to-end Gen-AI for industrial safety — both image generation and real-time video analysis pipelines.
  • A customizable image generator on HuggingFace's FLUX model with both GUI and headless (non-GUI) modes.
  • A logging framework capturing prompt and configuration data for full traceability and reproducibility.
  • Multi-threaded video analysis with Qwen VLM, OpenCV and Streamlit for real-time monitoring.

Results

Real-timeLive video analysis
2 modesGUI + headless generator
Full tracePrompt & config logging

Challenges

  • Latency vs. accuracy: VLM inference is heavy; multi-threading the frame pipeline kept analysis real-time without dropping critical frames.
  • Data scarcity: real hazards are rare, so FLUX-generated synthetic scenarios filled the gap with controllable, labeled edge cases.
  • Reproducibility: generative output is sensitive to prompts/config — the logging framework made every result traceable.

Learnings

  • Synthetic data is a force multiplier when real edge cases are scarce or dangerous to capture.
  • VLMs unlock open-vocabulary monitoring — you describe hazards in language instead of training a fixed-class detector.
  • For real-time CV, the systems engineering (threading, I/O) matters as much as the model.
Next project
Diabetic Foot Ulcer Detection