AI video analytics starts with the camera scene. A model can only analyze what the camera actually captures, so the site survey should happen before hardware selection and before a project promises a detection result.
Define the event in ordinary operational terms
Write down the visible condition that matters. It might be a worker entering a restricted area without PPE, a truck remaining in a fire lane, an e-bike entering an elevator lobby, material accumulating at a conveyor transfer point, or smoke appearing in a defined outdoor yard. Include the location, target, trigger, duration, excluded normal activity, required evidence, and response owner.
Check whether the target is visible
Review the camera height, viewing angle, distance, target pixel size, lighting, backlight, motion blur, occlusion, lens condition, and daytime and nighttime variation. A helmet or phone that is too small in the image cannot be made reliable by changing a threshold. The appropriate solution may be a new lens, a closer camera point, different illumination, or a different deployment architecture.
Review the existing video system
For a retrofit, confirm camera and NVR access, RTSP or ONVIF availability, stream encoding, resolution, frame rate, network location, stable long-running playback, available storage, and platform interface requirements. A sample stream from each camera group is usually needed because two cameras with the same catalog specification can produce very different images.
Set the rule for each point
Each camera should have a defined detection region, schedule, target filter, duration, confidence setting, alarm interval, and event receiver. These controls prevent one broad rule from generating alerts for normal work in another part of the image. They also create an operating record for later tuning.
Test with separate real-world video
Acceptance should cover normal and abnormal activity, difficult lighting, partial occlusion, different shifts, and other site-specific conditions. Use clips that were not used to create a custom model. The site team should agree whether each event is correct, what evidence is required, and how false alerts or missed events will be reported and adjusted.
Keep the human review loop
Video AI can reduce the amount of routine viewing, but it does not replace the person responsible for safety, operations, or property management. A useful project has a clear path from event to review, action, and feedback. The feedback examples then help refine regions, duration rules, thresholds, or a custom model.