Preview, detection and recording may share input or decoding stages, or use different resolutions, sampling rates, encoders and buffering policies. Do not assume identical processing paths. Adding individual workload figures is insufficient evidence of verified combined capacity.
Place counters along the actual data flow
| Stage | Records needed | Distinction to preserve |
|---|---|---|
| Input and decoding | Reception, valid frames, decode errors and waiting | Received data versus valid images |
| Detection processing | Scheduled, completed, skipped and result times | Intentional sampling versus unintended loss |
| Preview encoding and distribution | Encoded output, segments, reads and clients | Server output versus client display |
| Recording writes | Bytes written, files, indexes and errors | Temporary buffering versus confirmed storage |
| Playback reads | Retrieval, seeking, decoding and read times | Single-channel reading versus concurrent playback |
Define timing points and clock sources
Measure time at reception, queue entry, processing completion, segment availability and video display on the client. Define the start, end and clock source for each metric. Without a camera-side timestamp or an independent timing reference, measurements on the host computer alone cannot establish the time elapsed from camera exposure to on-screen video display.
Increase combined workloads gradually
Keep video modes and model versions unchanged during comparisons. Record resources, errors and waiting for preview-only, detection-only, recording-only and agreed combinations. After adding channels, clients, playback or retention tasks, check for bottlenecks in decoding, memory, network or disks. Keep unmeasured combinations marked as unverified.
Preserve recovery evidence
After input loss, reconnection, insufficient disk space or a service restart, check states and actual outputs for each function. A live service process is insufficient evidence of continuous recording or valid results. Retain configurations, original counters, logs and reproduction steps. This article promises no universal concurrent-channel capacity.
Related services: edge AI deployment and optimization and project enquiries.