Published: 2026-08-13 | English reference version of the Chinese source article.
What the report covers
This page summarizes the visual algorithm recognition test report issued by the Software Evaluation Center of the School of Computer Science and Technology, University of Science and Technology of China, for Hefei Xinhuo Information Technology Co., Ltd. The report number is XH-AI-VAL-20260809-V2 and the report date is August 9, 2026.
The evaluation covered restricted-area intrusion, helmet use, reflective vest use, smoking, phone use, smoke and flame, illegal parking and license plate recognition. It used a Rockchip NPU for local single-image inference with a batch size of one. The output form includes detection boxes, class labels, confidence values and structured alarm results.
Sample scope and measurement method
Each algorithm category used 3,000 single-image test samples, for 24,000 images in total. The samples include regular scenes, alarm or abnormal scenes, night and low-light scenes, dense occlusion, long-distance small targets and negative samples that can cause confusion.
The report defines recognition rate as the proportion of existing targets for which the class and state are judged correctly. Misrecognition refers to an incorrect class or state after a target is detected. False alarm rate records non-target or non-alarm images that are output as targets or alarms. Single-image inference latency measures model inference and structured-result generation; it excludes video decoding, network transfer and alarm upload.
Published metrics for the eight evaluated tasks
The figures below are reproduced from the report's metric table. They are tied to the report's sample definition, inference setting and evaluation method. A live project still needs site acceptance against its own camera view, lighting and alarm rules.
| Scenario | Recognition rate | Single-image latency | Misrecognition limit | False-alarm limit |
|---|---|---|---|---|
| Restricted-area intrusion | at least 97.2% | no more than 28 ms | no more than 1.1% | no more than 0.8% |
| Helmet | at least 96.8% | no more than 24 ms | no more than 1.2% | no more than 0.9% |
| Reflective vest | at least 95.6% | no more than 18 ms | no more than 2.1% | no more than 1.5% |
| Smoking | at least 94.8% | no more than 18 ms | no more than 8.9% | no more than 3.8% |
| Phone use | at least 95.4% | no more than 27 ms | no more than 5.8% | no more than 3.2% |
| Smoke and flame | at least 96.6% | no more than 31 ms | no more than 2.6% | no more than 1.8% |
| Illegal parking | at least 95.1% | no more than 33 ms | no more than 1.3% | no more than 1.0% |
| License plate recognition | at least 97.5% | no more than 22 ms | no more than 0.9% | no more than 0.6% |
Why the site test still matters
An image test checks model behaviour under a defined sample set. A delivered system also depends on camera optics, target size, video compression, lighting, scene rules, event deduplication and the interface that receives the alarm. Those factors must be checked with the customer's live streams.
For that reason, project acceptance should retain the test boundary, camera position, model version, parameter version, false alarms, missed events, captured evidence and delivery result. This makes later tuning traceable instead of treating every issue as a generic accuracy problem.