Skip to content
Documentation
Part IX: High-Scale System Design Case Studies & Applied Infrastructure
Chapter 34: Media Streaming & Cloud Storage Systems
34.3 System Design: Real-Time CCTV Video Analytics & Theft Detection Engine (RTSP Stream Ingestion, Edge-Cloud Hybrid Inference, Motion Detection Filters, Object ReID, Spatial-Temporal Action Recognition, Evidence Clip Archival)

34.3 System Design: Real-Time CCTV Video Analytics & Theft Detection Engine (RTSP Stream Ingestion, Edge-Cloud Hybrid Inference, Motion Detection Filters, Object ReID, Spatial-Temporal Action Recognition, Evidence Clip Archival)

What it is

A real-time CCTV video analytics and theft detection engine converts live camera streams into privacy-aware detections, tracked identities, spatial-temporal events, and reviewable evidence. Its mental model is a filter cascade: cheap motion gates reduce work, models produce observations, tracking and action recognition establish context, and an evidence service preserves only justified clips.

How it works

RTSP or another camera protocol provides a live media stream to an edge gateway. The gateway authenticates the camera, normalizes timestamps, drops malformed frames, and maintains bounded per-camera queues. A motion filter or scene-change detector removes static footage before expensive inference. Selected frames go to a detector, and detections enter a tracker that assigns short-lived object identities across nearby frames.

    flowchart LR
    A[Camera RTSP stream] --> B[Edge gateway]
    B --> C[Motion and quality filters]
    C --> D[Local object detector]
    D --> E[Cloud detector fallback]
    D --> F[Tracker and object ReID]
    E --> F
    F --> G[Spatial-temporal action model]
    G --> H[Event policy and confidence gate]
    H --> I[Evidence clip service]
    H --> J[Operator notification]

The edge-cloud split follows latency, bandwidth, and privacy. Edge inference handles motion filters and a compact detector when the camera link is intermittent or the scene is privacy-sensitive. The cloud receives selected crops, embeddings, or events for larger models, cross-camera re-identification, and fleet-wide model management. Uploads should carry only the data needed for the next decision, and the policy should record whether a frame was processed locally or remotely.

    sequenceDiagram
    participant C as Camera
    participant E as Edge gateway
    participant M as Motion filter
    participant D as Detector
    participant T as Tracker and ReID
    participant A as Action recognizer
    participant V as Evidence service
    C->>E: RTSP frames with timestamps
    E->>M: Sampled frames
    M-->>E: Candidate frames only
    E->>D: Local or remote detection request
    D-->>T: Boxes, classes, scores
    T->>A: Tracked identities and motion tracklets
    A-->>V: Event, confidence, camera references
    V-->>E: Clip retention acknowledgement

Re-identification, or ReID, matches an observation to a tracklet across cameras or time gaps. Appearance embeddings can help, but lighting, occlusion, uniforms, and similar-looking people create false matches. A spatial-temporal action recognizer combines a person track, a region, and a time interval to distinguish a suspicious action from ordinary movement. Its output is evidence for a policy engine, not proof by itself; thresholds, calibration, human review, and an audit trail matter.

Evidence archival should be selective. The service can retain a rolling buffer at the edge, fetch a pre-event window and post-event window after a confirmed event, encrypt the clip, and attach a manifest with camera, time, model version, policy version, and evidence hash. Access is time-bound and role-scoped. A privacy-preserving deployment masks faces for routine review and keeps an unmasked derivative only when an authorized investigation requires it.

retention:
  rolling_buffer: 30s
  pre_event_window: 15s
  post_event_window: 45s
  event_clip: 90d
  access_log: 2y
privacy:
  face_masking: routine_review
  unmasked_export: investigation_role
  regions_excluded: [lobby_restroom]

Tradeoffs

  • Always-on recording — preserves evidence, but raises storage, privacy, and review costs even when no event occurs.
  • Event-triggered recording — reduces retained data, but may miss context if the trigger runs late or the camera drops frames.
  • Edge-only inference — reduces bandwidth and keeps more processing local, but limits model size, fleet updates, and cross-camera context.
  • Cloud inference — provides stronger models and fleet-wide coordination, but increases bandwidth exposure, latency, and data-residency obligations.
  • Motion filtering first — reduces inference cost, but motion can be caused by lighting or weather while a theft can occur in a low-motion region.
  • Higher-recall filtering — catches more candidate events, but increases compute and false positives that can overwhelm operators.
  • Online tracking — supports immediate alerts, but tracker resets and occlusion can fragment an identity.
  • ReID across cameras — connects activity, but introduces privacy risk and false associations between similar people.
  • Fixed confidence threshold — is easy to operate, but can behave differently across camera quality, weather, and site populations.
  • Human-reviewed escalation — improves safety and accountability, but adds response time and requires trained operators.
  • Immutable evidence retention — supports investigations, but conflicts with minimization, deletion, and regional privacy requirements unless retention is bounded.
  • Masked review media — reduces bystander exposure, but can make evidence less useful for identification; unmasked access needs strong authorization and auditing.

When to use

You need edge filtering when camera links cannot carry continuous full-resolution video to a cloud region.

You need event-triggered recording when retention policy prohibits or cannot justify storing every frame.

You need ReID when an investigation must connect a track across camera boundaries and the consent or legal basis permits it.

You need spatial-temporal action recognition when a single-frame object label is insufficient to explain an event.

You need evidence archival with audit logs when detections can affect a person, employee, or customer.

Alternatives

Cloud video management — centralizes operations and retention, but depends on continuous connectivity and can move sensitive footage across regions.

Edge recording appliance — improves local availability and control, but requires field maintenance, storage sizing, and remote fleet management.

Motion-only detection — is inexpensive and easy to tune, but misses activity that does not create a large image difference.

Manual review — provides human context and avoids trusting a model alone, but does not scale continuously to many cameras.

On-device biometric recognition — can reduce round trips, but creates stronger consent, retention, and tamper-resistance requirements.

Related

34.2 System Design: Google Drive / Cloud Sync Architecture (Chunking Engine, File Delta Sync, Metadata Storage, Block Deduplication)Chapter 34 References