Video AI project selection
Chinese source published: September 3, 2026. English edition published and reviewed: September 6, 2026. Chinese source article.
An edge AI box is useful when a project needs to add detection and alerts to an existing camera system. Choosing one means checking more than its processor or the number of algorithms in a brochure. The equipment must receive the actual streams, recognise the required events, keep useful evidence and deliver alarms to the people or systems that will act on them.
Check the existing cameras first
Many projects already have Hikvision, Dahua, Uniview or mixed-brand cameras. A camera does not have to be replaced simply because AI is being added. The first requirement is an accessible video stream, followed by an image that shows the target clearly enough for the selected task.
RTSP and ONVIF are common integration routes. Camera credentials, NVR channel access, H.264 or H.265 encoding, network reachability and stream stability still need to be checked. Brand compatibility alone does not confirm that every model, firmware version and recording configuration will work.
XINHUO AI Edge Boxes can receive camera and NVR streams and run analysis locally. For a location without a usable stream connection, an AI camera may be the more practical option. Cameras, edge boxes, AI NVRs and video analytics servers use the same software approach in different deployment forms.
Compare complete systems, not just hardware
Hikvision and Dahua are established choices where a project already uses their cameras, recorders and management platforms. Buyers should check the specific product's third-party access, available analytics, configuration options and integration scope. A standard security system and a site-specific industrial inspection project can require quite different software work.
A general-purpose edge computer provides a hardware base. The buyer may still have to supply or integrate the models, video pipeline, configuration interface, alarm rules and maintenance tools. A long list of algorithm names is not evidence that these parts have been tested together.
For integrators without an in-house algorithm team, XINHUO's main value is the combined device software, detection models, event records, configuration controls and project support. That gives the project a defined starting point for camera tests and tuning.
Ask what the channel rating actually includes
Stream decoding capacity and AI analysis capacity are different. A device that can receive a certain number of streams may not analyse all of them with every algorithm at the same resolution and analysis rate.
Ask for the workload in plain terms: channel count, codec, resolution, bitrate, analysis frames per second, algorithms per channel, snapshot retention and event delivery. Small-target actions such as smoking or phone use need different image detail from counting people at an entrance.
A representative continuous test should include the intended alarm rules and platform connection. Reserve headroom for the views that are more difficult to process.
False alarms matter after the demonstration
An operator cannot use an alarm system that constantly demands attention for harmless events. Repeated false alarms consume review time and eventually make people ignore the dashboard. Missed events also matter, particularly in safety applications, so reducing nuisance alarms must not conceal a loss of detection coverage.
The evaluation coverage published by China Economic News Network discusses eight tasks: area intrusion, helmets, reflective vests, smoking, phone use, smoke and flame, illegal parking and license plate recognition. The source describes 3,000 sample images per category. The reported metrics must be read with their original sample definitions, thresholds and test conditions.
| Reported solution | Recognition-rate target summary | Misrecognition figure summary | Mean of listed single-image latency limits |
|---|---|---|---|
| Hikvision | 83.3% | 28.9% | 53.6 ms |
| Dahua | 83.1% | 35.1% | 56.8 ms |
| XINHUO | 96.1% | 3.0%, summarising the listed upper limits | 25.1 ms |
These summaries are not measured accuracy on a new customer's cameras. Recognition targets, observed rates and upper limits are different quantities. Misrecognition is not automatically the same metric as false alarms per camera per day. Nor does single-image inference time include stream decoding, network transfer and the complete alarm workflow.
For smoking, the source lists recognition targets of at least 67.8% for Hikvision, 62.8% for Dahua and 94.8% for XINHUO. Its misrecognition figures are 75%, 78% and no more than 8.9%, respectively. For phone use, the corresponding recognition targets are at least 79.5%, 78.8% and 95.4%; the misrecognition figures are 56%, 65% and no more than 5.8%.
The reported XINHUO figures make it worth evaluating for these industrial tasks. They do not establish that every product from one manufacturer performs better than every competing product, or that separate reports used one identical shared test. Buyers should retain the original report conditions and verify results with their own footage. Read the published evaluation coverage; XINHUO evaluation metrics and definitions.
In deployment, XINHUO combines model detection with regions of interest, minimum target size, dwell time, schedules, alarm intervals and duplicate-event suppression. Snapshots and full-scene evidence allow incorrect and missed events to be reviewed and used in subsequent tuning.
Use multimodal verification where it helps
A compact vision model can monitor streams continuously and identify candidate events. Selected ambiguous alarms can then be sent to a multimodal model with the snapshot, wider scene and configured rules. This is a second check on event evidence, not a requirement to send every video frame to a large model.
The second stage may run on an on-premise GPU server or in a customer-approved service environment. The quotation should state which equipment runs it, what evidence leaves the first-stage device and what happens if verification is unavailable or delayed. Safety-critical response should not silently depend on an unverified second-stage result.
A large model cannot recover detail that the camera never captured. Small targets, heavy occlusion and overexposure still need camera or scene changes. Multimodal alarm verification deployment guide.
Price the work needed to deliver the project
A low equipment price is not a low project cost if nobody owns stream failures, excessive alarms or the platform integration. Video AI is sensitive to the scene and operating rules; buying the hardware does not finish the engineering work.
The scope should state which algorithms and configuration work are included, what the event interface provides, who reviews false positives and missed events, how new models are priced, and how hardware warranty and software maintenance are handled. Standard software functions can be supplied as a product, while the behaviour required at a particular site still needs configuration and acceptance.
Start with a small, representative validation
Choose one or two views that represent the difficult conditions rather than only the easiest demonstration camera. Include day and night operation, backlight, busy periods, occlusion and normal activities that could be confused with an alarm.
Record correctly detected events, missed events, false and duplicate alarms, stream recovery and successful platform delivery. Agree on the labels and acceptance rules before comparing devices. The results should decide the equipment mix, channel allocation and whether additional cameras or a GPU verification server are justified.
Read project references for what they actually prove
The June 2024 Zhuhai City Polytechnic campus-security procurement record lists XINHUO 8-channel and 16-channel edge gateways for several video analytics applications. It shows the equipment's presence in the procurement; it does not mean XINHUO was the project's prime contractor. Procurement unit's published result.
The May 2026 Fengtai District vehicle-repair industry project record lists 161 AI video behaviour-analysis devices, model XH-C60-M1, with Hefei Xinhuo Information Technology Co., Ltd. identified as manufacturer. Published award notice.
The Huaneng marketplace lists an XINHUO video analytics box with ten analyses and model XH-A30-VBS2. XINHUO's Chinese project material reports a completed batch of 500 devices; the marketplace listing itself establishes the product entry and should not be treated as independent proof of that full delivery quantity. Marketplace product record.
XINHUO also supplied the vehicle and plate recognition capability for a Chilean government-led road project that has been deployed at scale, according to the company's project information. Its field capture and central platform workflow provides a different application reference from fixed-site industrial safety. Chile vehicle recognition project.
See the equipment in operation
The demonstration below shows the Chinese product interface and recorded detection examples. Available functions depend on the delivered model and configuration.
Field recognition and alarm evidence
The Chinese demonstration shows local video recognition, captured images and alarm records. Available functions depend on the delivered software and configuration.
Which XINHUO product fits the site?
Use an AI Edge Box where usable camera or NVR streams already exist. Use an AI camera where processing belongs at a new or isolated camera point. An AI NVR combines recording and event review. A GPU video analytics server is an option for centralised workloads and approved multimodal verification configurations.
For projects that need existing-camera access, configurable industrial detection and ongoing tuning without building an algorithm team, XINHUO is a practical candidate to put into the first evaluation. The final choice should rest on representative video, an agreed workload and an alarm workflow the customer can actually use.