Where AI runs affects bandwidth, cost, outage behavior and future flexibility. Choose the architecture around the task and operating responsibilities.
Compare the processing locations
| Architecture | Useful characteristics | Dependencies to examine |
|---|---|---|
| Edge: in the camera | Distributes processing across supported cameras and can generate events near the source. | Camera processor, supported application, firmware, licenses and metadata integration. |
| Local server or appliance | Can add processing to compatible streams and centralize selected analytics. | Compute capacity, stream decoding, maintenance, resilience and competition with recording workloads. |
| Cloud service | Can centralize remote search and service management across locations. | Connectivity, upload/indexing behavior, subscriptions, retention and account access. |
| Hybrid | Combines local recording or inference with remote management or search. | Precisely which functions depend on each component; “hybrid” alone explains little. |
Draw four locations on the design
- Where is the original video recorded?
- Where is inference performed?
- Where are metadata and search indexes stored?
- Where do users authenticate, review events and change rules?
These locations may differ. A local recording system can depend on cloud authentication for a remote workflow. A cloud-managed service may continue some local operations during an interruption. Request the actual product behavior rather than assuming “cloud” means no local storage or “edge” means no service dependency.
Can existing cameras gain AI?
Sometimes a supported local or cloud analytics product can analyze existing streams. Reuse still depends on image quality, stream format, frame rate, authentication, licensing and the receiving platform. An older camera may produce an acceptable stream but lack the detail needed for the intended task.
Inventory each camera’s exact model, firmware and scene. Group devices by integration and image requirements. Pilot representative cameras before approving reuse of the entire inventory. Price exceptions and replacement work explicitly.
Open interfaces are only part of portability
ONVIF conformance applies to listed products and versions. Check the relevant profile on both sides and demonstrate the required functions. Analytics metadata, search indexes, model configurations and software licenses do not automatically transfer when hardware does.
| Layer | Owner question |
|---|---|
| Camera | Can another supported platform authenticate and record the required stream? |
| Analytics | Does the next platform consume the events and attributes we need? |
| Data | Can the owner export original recordings and necessary event information? |
| Licensing | Which licenses terminate, transfer or need replacement? |
| Service | Can another authorized provider administer this platform without replacing it? |
Outage tests to request
Consider internet loss, unavailable analytics processing, recording-server failure and loss of a network segment separately. Record which detections, searches, alerts, recordings and account changes continue. Coordinate any live test with the responsible parties so occupied operations and life-safety functions are not disrupted.
Ownership-cost comparison
Include cameras, processing appliances, servers, storage, licenses, cloud subscriptions, network changes, maintenance, updates, operator review time and future migration. A lower device price can be offset by additional processing or recurring services. A higher upfront design may still demand staff expertise and maintenance.
Fictional decision: a business retains usable cameras and adds an analytics appliance. It saves camera replacement cost but still needs licensing, compute support and an export plan. Evaluate the whole change rather than treating reused hardware as a free AI upgrade.
Continue with the manufacturer comparison, integration and ownership and the evaluation workbook.