Real-Time Scene Understanding#
Real-Time Scene Understanding deploys AI-powered live video analysis for RTSP streams with FFMpeg pythonic bindings, PyAV and OpenVINO Vision Language Models (VLMs). You can process streaming video, run prompt-driven scene understanding, and monitor both system and inference performance from a web dashboard.
The key features are:
VLM Model Support: Configure separate alert and deep-analysis OpenVINO VLMs through environment settings, with model files mounted from ov_models/.
Real-Time Streaming: WebRTC-based low-latency video preview for live monitoring workflows.
Prompt-Driven Scene Intelligence: Run natural-language prompts against live frames for flexible scene interpretation.
Performance Metrics: Live charts for CPU/GPU/RAM plus inference metrics such as TTFT, TPOT, and throughput.
Modular Architecture: Containerized application, streaming, metrics, and object-storage services.
Alert Mode: Alert-oriented output styling for binary prompt responses (for example, “Yes”/”No”).
Use Cases#
Real-Time Monitoring: Analyze security cameras, industrial lines, and public infrastructure with prompt-driven scene understanding.
Operational Safety: Detect safety-relevant conditions through targeted prompts in manufacturing, warehousing, and logistics environments.
Accessibility and Summarization: Generate concise, contextual scene descriptions to improve accessibility and operator awareness.
Performance Benchmarking: Compare VLM behavior across Intel hardware by tracking latency, throughput, and resource utilization.