How us107/capstone-project-25 works
Project type and purpose
• This is a full-stack, AI-powered video surveillance system. It ingests live or recorded video, processes it with computer-vision and deep-learning models, stores activity logs and relationships, and exposes results via interactive dashboards and reports.
File structure
• Top-level Python scripts each implement one stage of the pipeline:
• csvfiles_storage.py / newcloudcsvstorage.py – local and cloud CSV storage modules
• dashboardneo.py / realtime_dashboardneo.py / realtimeclouddashboard.py – Streamlit frontends for local and cloud dashboards backed by Neo4j
• extracted_frames.py – frame-extraction (FFmpeg/OpenCV)
• deepsort_yolo.py – object detection (YOLOv8) + tracking (DeepSORT)
• r3d_18activity.py – spatiotemporal action recognition (3D ResNet-18)
• generate_alerts.py – classify and emit real-time alerts
• generate_summary.py – compile summaries, invoke Groq LLM for PDF reports
• data_splitting.py – training/validation data prep
• testing.py – unit/integration tests
• requirements.txt and README.md document dependencies and high-level architecture.
README for additional insights
• Confirms real-time anomaly detection, centralized storage, interactive Streamlit dashboards, Neo4j for event-relationship graph, and Groq LLM for report generation.
Main components and their interactions
a. Video Ingestion
• sources: CCTV streams or video files
• script: extracted_frames.py
b. Object Tracking
• deepsort_yolo.py uses YOLOv8 + DeepSORT
c. Action Recognition
• r3d_18activity.py (3D ResNet-18)
d. Anomaly & Alert Generation
• generate_alerts.py
e. Summary & Report Generation
• generate_summary.py → Groq LLM
f. Storage
• csvfiles_storage.py (local CSV)
• newcloudcsvstorage.py (cloud CSV)
• Neo4j (graph storage of event relationships)
g. Dashboards / UI
• dashboardneo.py (local Streamlit + Neo4j)
• realtime_dashboardneo.py / realtimeclouddashboard.py for live updates
Architectural patterns & design principles
• Modular data-processing pipeline: each script is a standalone stage.
• Event-driven triggers or scheduler to chain scripts.
• Separation of concerns: compute modules vs. storage modules vs. UI modules.
• Extensibility hooks (e.g., Whisper for audio, edge deployment).
Technologies, frameworks, libraries
• Python ecosystem: OpenCV, FFmpeg, PyTorch
• Detection & tracking: YOLOv8, DeepSORT
• Recognition: 3D ResNet-18
• Dashboard: Streamlit
• Graph DB: Neo4j
• LLM: Groq (PDF report generation)