Updated: August 27, 2026
Can existing CCTV cameras be used?
Yes, when the project can obtain a usable RTSP stream, ONVIF access, GB28181 access or an NVR channel stream. The team still needs to check camera credentials, network reachability, codec, stream stability, image resolution and whether the target is visible enough for the intended algorithm.
When is an AI camera a better fit than an AI Edge Box?
An AI camera suits a new camera point or a distributed site where there is no practical way to obtain a network video stream. It can analyze video at the camera and send event data through 4G, a private network or another available connection. An AI Edge Box suits existing IP cameras and NVR streams that can be reached on the local network.
Does local AI video analytics require continuous cloud connectivity?
The real-time recognition, capture and local event logic can run on the local device. Internet connectivity is only needed when a project chooses a cloud service, remote platform access or a cloud-based review workflow. Local or isolated networks can retain events and synchronize them when an approved connection is available.
What does an 8-, 16- or 32-channel capacity mean?
Channel capacity is a planning range, not a fixed promise for every video task. Resolution, codec, frame rate, the number of algorithms on a camera, target density, live preview, event images and delivery frequency all use computing resources. The final channel mix should be checked with representative streams and a continuous run.
Which AI tasks can be configured?
The common catalogue covers PPE, smoke and flame, restricted-area entry, line crossing, parking and vehicle events, license plates, smoking, phone use, duty-post review, fire-lane occupation, camera availability, equipment states and conveyor risks. Each project should select only the events that can be seen clearly and have an agreed response rule.
Can a model be adapted for a site-specific object or rule?
Yes. XINHUOAI training can use representative local images or video frames for data preparation, annotation, training, evaluation and export to supported cameras, Edge Boxes, NVRs or servers. A new model must still be checked on the target device and with the live camera view before acceptance.
Why can false alarms occur after a model has been deployed?
Alarm quality is affected by camera angle, target pixels, backlight, shadows, night light, motion blur, rain, haze, occlusion and similar-looking objects. The system should retain event evidence and allow region, dwell-time, time-window, repetition and confidence settings to be reviewed against both normal and abnormal footage.
What is multimodal second-stage alarm verification?
A local detector can identify a candidate event and save a crop, full view, timestamp and rule context. For selected uncertain or high-impact events, a multimodal model can examine that evidence again. This reduces the number of alerts sent for manual review in suitable deployments. It needs a defined review model and computing path, such as an approved service or a local GPU server, and it does not remove the need for a well-positioned camera or clear event rules.
How can the AI events reach an existing platform?
Projects can receive structured event records through HTTP or MQTT, with the agreed fields for device, channel, time, algorithm, event type, confidence and evidence. The published API reference explains the event-message model. Existing interfaces and field mappings should be tested before the site goes live.
What should be prepared before an online demonstration or proposal?
A short description of the work area, camera count, camera or NVR brand, stream-access method, resolution, network topology, target event, expected response and a few representative screenshots or clips is usually enough to start. It prevents a demonstration from turning into a generic product presentation that does not match the site.