Direct answer
Start from the shortest event the system must detect and the tolerated response delay. Sample enough frames to observe that event reliably, then validate misses and compute cost on real footage.
Why this happens
A 30 fps camera does not require 30 inferences per second. Excess sampling wastes decode and accelerator capacity, while sparse sampling can miss fast or brief events.
Size the ingest path for the inference job, then keep the original stream, derived metadata and rendered overlays distinct.
A controlled test
Collect representative day, night and motion footage and replay it at several sampling rates with the same model threshold.
Change one variable at a time. Keep the camera model, firmware, endpoint and account recorded; then test network reachability, protocol response, media transport and decoding as separate layers.
Use a dedicated view-only account and a trusted local diagnostic tool. Redact credentials, private addresses and identifying data before sharing output.
Diagnostic sequence
| Check | Action | Evidence of progress |
|---|---|---|
| Event | Measure how long the target remains visible. | A minimum observation window is known. |
| Rate | Test several inference intervals. | Miss rate and delay are measured. |
| Compute | Track decode and accelerator utilization. | The selected rate fits peak capacity. |
| Fallback | Define what happens when the AI queue lags. | Frames are dropped or delayed by policy, not accident. |
Evidence to keep
Report event duration, sampling rate, detection delay, miss rate and compute usage together.
Boundary and safety note
Do not treat an AI alert as a final safety, employment, access or emergency decision without the approved human process.
For remote viewing, use a managed VPN instead of exposing RTSP or camera administration ports directly to the public internet.
SmartRTSP
SmartRTSP is a camera-focused RTSP and ONVIF viewer for Apple devices, Windows and Android. It fits direct viewing, discovery and multi-camera checks; keep a dedicated NVR or VMS when continuous recording, evidence export or centralized enterprise controls are required.
Frequently asked questions
Should AI process every RTSP frame?
Usually not; the required rate comes from event duration, motion and response targets.
Can I lower camera frame rate instead?
Sometimes, but recording or human viewing may need a different rate; separate the inference sampling policy from the source profile where possible.
What happens when inference falls behind?
Choose explicitly whether to drop old frames, reduce sampling or scale capacity so decisions are not based on stale video.
Primary references
- NVIDIA — DeepStream architecture overview
- ONVIF — Profile M for metadata and analytics
- IETF RFC 7826 — Real-Time Streaming Protocol 2.0
Related SmartRTSP guide
Open related guideMatch inference frequency to object speed, field of view and response time instead of decoding every frame by habit.