CrowdDNA: A Crowd Risk Prediction System Based on Graph Neural Networks and Computer Vision

CrowdDNA is an intelligent platform combining YOLOv8, ByteTrack, Graph Attention Network (GAT) and GRU temporal models to perform real-time risk classification on crowd videos. Developed by Piyush Gupta (AI & Data Lead) and Aayushi Gupta (Pipeline & Frontend Lead), it models crowd spatial relationships and kinematic features via dynamic graphs to predict local risks (Safe/Congesting/Critical). Hosted on GitHub and Hugging Face Spaces, it addresses limitations of traditional crowd monitoring.

Traditional video surveillance in crowded places (large events, transport hubs) only provides visual feeds but fails to actively identify potential hazards like stampedes. CrowdDNA fills this gap by integrating computer vision and graph neural networks to 'understand' crowd behavior patterns, enabling early risk warnings.

1. Video Preprocessing: Validate MP4/AVI files and extract key frames.
2. Detection & Tracking: Use YOLOv8n (lightweight) for pedestrian detection and ByteTrack for persistent multi-object tracking.
3. Dynamic Graph Construction: Build graphs based on spatial proximity, relative speed, and density gradient.
4. Risk Classification: Combine GAT (graph processing) and GRU (temporal modeling) to output 3 risk levels; PyTorch model exportable to ONNX.
5. UI: Gradio-based interface on Hugging Face Spaces, providing annotated videos and risk timelines.

1. Semantic Behavior Understanding: Beyond counting, captures crowd behavior semantics.
2. Dynamic Graph Modeling: Treats crowds as dynamic graphs instead of individual sets.
3. End-to-End Pipeline: Covers video ingestion to risk visualization.
4. Lightweight: Uses YOLOv8n and ONNX for balance of accuracy and inference speed.

• Public Safety: Real-time monitoring of stadiums, concerts, religious gatherings.
• Traffic Management: Crowd flow management in subway stations, airports.
• Urban Planning: Analyze historical data to optimize public facility layout.
• Emergency Response: Provide quantitative risk indicators for security decision-making.

The project uses a two-person collaboration model with clear standards:

• AGENTS.md: AI-assisted development workflows.
• WORKFLOW.md: Git branch strategies and code submission processes.
• CODING_STANDARDS.md: Unified code style and module interfaces. This approach is a reference for open-source projects.

CrowdDNA demonstrates GNN application to social safety, with a reusable pipeline (detection→tracking→graph building→spatiotemporal modeling). Future directions: finer-grained behavior recognition (panic/fights) and lower latency via multi-modal models and edge computing.