Start with the IP cameras or NVR you already have and verify that they expose an RTSP or ONVIF stream. Add an always-on local box with storage: a mini PC is one option. This guide describes a proposed DIY architecture, not a tested kit. Confirm stream access and compute requirements before buying hardware. A GPU or edge accelerator is optional, selected only after checking software compatibility and workload.
Build the pipeline in stages: RTSP ingest → local motion and person/vehicle detection → event snapshots → a vision-language model checking a plain-language rule → an alert with a snapshot. Package checks can be a separate snapshot task. Begin with one stream and one observable rule. Define a zone and the times when its rule applies. As proposed noise controls, group detections from the same visit into one event and add a cooldown before repeating an alert. Connect an authorized messaging-app or email integration for the DIY build; test delivery separately. Generate a daily summary from the event log, including uncertain checks for review.
Keep video and event storage on site, limit retention and restrict access. Prefer local snapshot analysis when the chosen model fits your hardware; any cloud snapshot analysis must be an explicit choice, with masking and a review of what leaves the site. Alerts with snapshots also share images with the chosen messaging provider and recipients. Obtain consent where required, provide signage, and check local law before recording or sharing images. Do not point cameras at neighbours or public space; use privacy masks as an additional safeguard. Review missed events and false positives manually. This is assistive monitoring and must not replace safety systems or human supervision.