VICSTA Hackathon - Grand Finale

VIT College, Kondhwa Campus | 5th – 6th March

---

Team Details

• Team Name: Team Spectrum
• Members:
  • Veer Gandhi
  • Sanchit Borikar
  • Purvesh Didpaye
  • Ashraf Ahmed
• Domain: Productivity & Security (PS-04)

---

Project

Problem: Over 20 million cameras globally record everything equally, wasting approximately Rs. 40,000/month per mid-sized deployment on storing empty, idle footage that no one ever watches. Furthermore, traditional market solutions rely on binary motion detection, which triggers false alarms from background shadows and remains completely blind to stillness (such as a loiterer).

Solution: EdgeVid LowBand is a neuromorphic edge-AI DVR that fundamentally changes how video is processed.

• SNN Spike Gate: Acts as a neuromorphic pre-filter, bypassing 80% of compute load by skipping idle frames entirely before object detection runs.
• Hardware-Accelerated YOLOv8: Runs precision object detection strictly on SNN spikes, grading every frame with an Intelligence Score (0–100).
• Dynamic ROI Compression: Applies extreme spatial compression to backgrounds (15% quality) while keeping the target subject crystal clear (88% quality), achieving a 70% overall reduction in storage overhead.
• Forensic Auditing: Features a predictive 30-second pre-buffer for anomaly events (like loitering) and logs all metrics to an immutable SQLite database without relying on external cloud APIs.

---

System Architecture

%%{init: {'theme': 'base', 'themeVariables': {'background': '#ffffff', 'mainBkg': '#ffffff', 'edgeLabelBackground': '#ffffff', 'lineColor': '#64748b'}}}%%
flowchart LR
    classDef cam      fill:#dbeafe,stroke:#2563eb,stroke-width:2px,color:#1e3a8a
    classDef snn      fill:#ede9fe,stroke:#7c3aed,stroke-width:2px,color:#3b0764
    classDef yolo     fill:#dbeafe,stroke:#1d4ed8,stroke-width:2px,color:#1e3a8a
    classDef score    fill:#ffedd5,stroke:#ea580c,stroke-width:2px,color:#7c2d12
    classDef skip     fill:#f1f5f9,stroke:#94a3b8,stroke-width:2px,color:#475569
    classDef alert    fill:#fee2e2,stroke:#dc2626,stroke-width:2px,color:#7f1d1d
    classDef compress fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d
    classDef infra    fill:#e0f2fe,stroke:#0284c7,stroke-width:2px,color:#0c4a6e

    subgraph CANVAS[" "]
        CAM["Camera Input\nWebcam / RTSP / IP"]:::cam
        PREBUF["Pre-Event Buffer\n30s Circular"]:::snn
        SNN["SNN Spike Gate\nSkips 80% idle frames"]:::snn
        SKIP["Skip Frame\nZero Compute"]:::skip
        YOLO["YOLOv8-nano\nObject Detection"]:::yolo
        ANOM["Anomaly Detector\nLoitering Detection"]:::alert
        SCORE["Frame Score\n0 to 100"]:::score
        DEC{{"Threshold"}}:::score
        HEAVY["Heavy Compress\nScore below 30\n15% JPEG"]:::skip
        ROI["ROI Compress\nScore above 60\nSubject 88% / BG 12%"]:::compress
        DB[("SQLite\nForensic Log")]:::infra
        API["FastAPI Server\nWebSocket"]:::infra
        DASH["React Dashboard\nLive Feed / Clips / Alerts"]:::compress

        CAM --> SNN
        CAM --> PREBUF
        SNN -->|"No Spike"| SKIP
        SNN -->|"Spike"| YOLO
        SNN -->|"Anomaly"| ANOM
        YOLO --> ANOM
        YOLO --> SCORE
        SCORE --> DEC
        DEC -->|"below 30"| HEAVY
        DEC -->|"above 60"| ROI
        PREBUF --> DB
        ANOM --> DB
        ANOM --> API
        ROI --> API
        HEAVY --> API
        DB -->|"Query"| API
        API --> DASH
    end

    style CANVAS fill:#ffffff,stroke:#e2e8f0,stroke-width:2px

Loading

---

Attribution

Library	Role	License
SpikingJelly	Neuromorphic SNN spike gate	Apache-2.0
YOLOv8-nano (Ultralytics)	Real-time object detection and frame scoring	AGPL-3.0
OpenCV	Camera capture, frame processing, ROI extraction	Apache-2.0
FastAPI	Async backend API and WebSocket server	MIT
React.js	Real-time surveillance dashboard	MIT
SQLite	Local forensic event database	Public Domain
zstandard (zstd)	High-speed lossless compression for EVENT frames	BSD
py7zr	LZMA2 batch archiving for IDLE frame sequences	LGPL-2.1