Nexus-safe

NEXUS-SAFE turns existing CCTV cameras into an intelligent security network that detects threats, fuses multi-camera context, scores risk, and delivers real-time actionable alerts at the edge.

Comment

Inspiration

Traditional CCTV systems mostly record what happened. They do not understand what is happening, connect events across cameras, or help security teams prioritize real threats.

We wanted to turn existing CCTV infrastructure into an intelligent security system that can detect abnormal activity in real time while remaining privacy-preserving and practical to run on ordinary edge hardware.

That led us to build NEXUS-SAFE: an edge-native, multi-camera video intelligence platform that combines computer vision, object tracking, spatial security rules, deterministic alert management, and contextual risk intelligence.

What it does

NEXUS-SAFE transforms multiple CCTV feeds into a real-time security intelligence network.

It can:

  • Detect vehicles using a fine-tuned YOLOv8n ONNX model.
  • Track objects using ByteTrack with anonymous camera-local IDs.
  • Detect restricted-zone intrusion, loitering, tripwire crossings, and wrong-way movement.
  • Process multiple camera streams concurrently using OS-level multiprocessing.
  • Maintain low-latency video processing using bounded queues and DROP_OLDEST backpressure.
  • Deduplicate repeated alerts and manage incidents through a deterministic NEW → ACKNOWLEDGED → RESOLVED lifecycle.
  • Fuse spatial, severity, persistence, repeated activity, and cross-camera context into an explainable 0–100 risk score.
  • Learn runtime traffic baselines and identify behavioral patterns.
  • Correlate related activity across cameras without attempting to identify the same person.
  • Provide operators with a real-time command center, incident triage, telemetry, and audit history.

The system is designed to run inference on-premise at the edge, with no facial recognition and no cross-camera biometric identification.

How we built it

The system is organized as a complete real-time pipeline.

Video ingestion: Four synchronized CCTV streams are captured using dedicated reader threads and bounded ring buffers. When processing becomes overloaded, obsolete frames are dropped instead of allowing latency to grow indefinitely.

Concurrent processing: Python multiprocessing workers provide true OS-level parallelism while controlling OpenCV and numerical-library thread counts to avoid CPU contention.

Computer vision: A fine-tuned YOLOv8n model is exported to ONNX and executed using ONNX Runtime. ByteTrack maintains camera-local anonymous tracking IDs.

Security rules: Detection trajectories are evaluated against spatial geofences, tripwires, loitering rules, and directional constraints.

Alert authority: Triggered events pass through deterministic deduplication, cooldown, and auto-resolution logic before being persisted in SQLite using WAL mode.

Contextual intelligence: A deterministic five-factor risk engine combines spatial risk, persistence, severity, repeated activity, and cross-camera correlation. A Welford online baseline learns normal activity from runtime data, while rolling temporal clusters connect related incidents.

Operator interface: A React 18 + Vite dashboard communicates with the backend through REST APIs and persistent WebSockets, providing live camera views, incident triage, system telemetry, and audit history.

Challenges we ran into

The hardest part was not simply detecting objects. It was building a system that remained reliable when multiple cameras were producing data simultaneously.

Naive video pipelines quickly accumulate stale frames and increasing latency under CPU pressure. We solved this using bounded queues with DROP_OLDEST backpressure and adaptive processing.

We also had to prevent duplicate alerts from overwhelming operators. Instead of generating independent notifications for every detection, NEXUS-SAFE uses deterministic alert suppression, cooldowns, and a persistent incident state machine.

Another challenge was contextual reasoning. Correlating activity across cameras without incorrectly claiming that two observations represent the same person required us to design temporal and category-based correlation rather than biometric identification.

Accomplishments that we're proud of

We built NEXUS-SAFE as a complete working system rather than a standalone AI model.

Our validation includes:

  • 111/111 backend tests passed
  • 16/16 Phase 6 intelligence tests passed
  • 1,604 frontend modules successfully built
  • 4 synchronized camera feeds
  • Approximately 21 FPS per camera on commodity CPU hardware
  • Approximately 23.4 ms inference time per frame
  • Deterministic duplicate-alert suppression
  • Persistent incident history with SQLite WAL
  • Real-time WebSocket operator dashboard
  • Edge-only inference with anonymous tracking

Most importantly, the system connects low-level computer vision with higher-level security reasoning and gives an operator an actionable view of what is happening.

What we learned

We learned that real-world AI systems require much more than model accuracy.

A useful security platform needs reliable ingestion, concurrency control, latency management, deterministic state handling, explainable decisions, persistence, observability, and a clear separation between authoritative security events and analytical insights.

We also learned that privacy should be part of the architecture from the beginning. NEXUS-SAFE deliberately avoids facial recognition, biometric identification, and cross-camera identity matching while still extracting useful security context.

What's next for NEXUS-SAFE

Our next goal is to expand NEXUS-SAFE from a strong multi-camera security platform into a broader edge intelligence system.

Future work includes supporting more object and threat categories, stronger adaptive anomaly detection, additional edge hardware acceleration, richer incident investigation tools, and integration with existing security infrastructure.

We also plan to improve the contextual intelligence layer so that operators receive not only alerts, but concise explanations of why an event is considered risky and what action may be appropriate.

The long-term vision is simple: make existing cameras intelligent enough to understand situations, not just record them.

Built With

Share this project:

Updates

Submission history