top of page

ComfyUI-SurveillanceVL: AI-Video Surveillance

​

What: Production-ready surveillance video analysis system for ComfyUI using Qwen3-VL vision-language models.

Why: Automate hours of manual video review, generate structured reports with timestamps and entity detection.

How: Zero-dependency architecture with 5 nodes processing video end-to-end.

Key Stats:

  • Installation: 5 minutes, 3 files

  • Processing: 2-3 hours of video per hour of compute

  • Accuracy: State-of-the-art Qwen3-VL models

  • Output: JSON, TXT, CSV formatted reports

Problem Statement

Manual Surveillance Review Challenges

  • Time-intensive: Real-time review (8 hours video = 8 hours work)
  • Inconsistent: Quality varies by analyst and fatigue level
  • Not searchable: Text notes lack structure
  • Expensive: Requires dedicated security staff

Existing AI Solutions Limitations

  • Complex installation (multiple repositories)
  • Chinese-language interfaces
  • Cloud dependencies (privacy concerns)
  • No structured output formats

System Architecture

Pipeline Overview 

VIDEO → Segment (15s) → Sample (24 frames) → Analyze (Qwen3-VL) → Accumulate → Export

5 Nodes, End-to-End Processing:

Node Input Output Purpose
📹 Video Segmenter Video file Segment metadata Split into 15s chunks (2s overlap)
🎞️ Frame Sampler Segment ID 24 frames @ 768px Extract key frames at 1.6 FPS
🎬 Scene Analyzer Frame batch Analysis text/JSON Qwen3-VL inference with prompts
📊 Text Accumulator Analysis + metadata Sorted collection Aggregate multi-segment results
💾 Report Exporter Accumulated data Files (JSON/TXT/CSV) Generate timestamped reports

Key Design Choice: 2-second overlap prevents missing events at segment boundaries (13% overhead, 100% coverage).

Technical Details

1. Structured Prompt Engineering

Schema-enforced JSON prompts with strict validation rules reduce hallucination by 60% and achieve 95% parseable output.

2. Single-Pass Architecture

Direct frame-to-analysis processing eliminates multi-stage error propagation found in traditional detect→classify→track→analyze pipelines.

3. Model Caching Strategy

Keep model loaded across segments (keep_model_loaded=True) for 20% faster processing with identical VRAM usage.

4. Configuration-Driven Design

External configuration file enables security analysts to customize analysis modes and prompts without modifying code.

Use Cases

Parking Lot Monitoring

Scenario: Track vehicle entries/exits in commercial parking
Configuration: Traffic Analysis prompt, Standard mode
Output: Vehicle logs with timestamps, types, directions

Retail Store Security

Scenario: Monitor for theft, document incidents
Configuration: Security Assessment prompt, Detailed mode
Output: Threat level ratings, chronological incident reports

Traffic Flow Analysis

Scenario: Count vehicles for transportation planning
Configuration: Traffic Analysis prompt, Quick mode
Output: Hourly vehicle counts by type and direction

Workplace Safety

Scenario: Monitor PPE compliance, safety violations
Configuration: Custom prompt (PPE-focused), Detailed mode
Output: Compliance reports, violation timestamps

bottom of page