Hikvision's Guanlan boosts video AI accuracy

1/ Hikvision said on Sept. 8 that it has upgraded its Guanlan AI applications, a video-analysis stack the company launched more than a year ago, with a pitch centered on higher accuracy, lower cost and more natural interaction for security and operations systems. (eqs-news.com) 2/ The company’s framing is straightforward: video systems should not just record scenes, but “understand” what they are seeing, know what has already been recorded, and respond to user requests through natural-language interfaces. That description came from Hikvision’s product announcement. (eqs-news.com) 3/ In practical terms, Hikvision is describing two linked shifts. The first is better real-time detection at the edge, where cameras and sensors try to filter noise before it reaches operators. The second is easier retrieval and analysis of archived footage, using text-based search rather than manual scrubbing through video. (eqs-news.com) 4/ The most concrete performance claim in the release is around perimeter protection. Hikvision said DeepinViewX cameras running the model cut false alarms by 90%, detect at twice the range of traditional cameras, and generate half as many repeated alarms on the same target. Those are company claims in the announcement; no independent validation was provided there. (eqs-news.com) 5/ That matters because perimeter systems are prone to nuisance triggers from animals, weather and moving vegetation. Hikvision’s argument is that lowering those false positives reduces manual review work and makes alarms more usable in live operations. (eqs-news.com) 6/ Hikvision also said Guanlan is being applied beyond generic object detection into scenario-specific analytics. The examples it gave were personal protective equipment detection on construction sites, fall detection in care homes, and smoking detection in gas stations. In those specialized settings, the company said applications “typically reduce false alarms by 70% to 90%.” (eqs-news.com) 7/ The broader product idea is an enterprise video AI stack rather than a single model. Hikvision’s Guanlan materials describe a family of large-scale AI models used across cameras, network video recorders and other AIoT systems, with deployment options spanning edge, center and center-edge fusion. (hikvision.com) 8/ That architecture helps explain the “lower cost” part of the pitch. Hikvision says some workloads can run at the edge for low-latency decisions, while others can be handled centrally to lower inference costs for intelligent applications. The company presents that as a way to match compute placement to the job rather than running everything in one place. (pro-av.hikvision.com) 9/ The “smarter interaction” part is mostly about multimodal search. Hikvision’s Guanlan product pages say its large multimodal models align images and text in a shared vector space, enabling searches for objects and events in video footage using natural language. That is the mechanism behind the company’s claim that systems can “know” what has already been recorded. (hikvision.com) 10/ Read narrowly, this is an enterprise surveillance and operations story, not a consumer AI launch. The target use cases in Hikvision’s own materials are industrial sites, construction, care facilities, gas stations and large perimeters where operators already manage multiple cameras and sensor feeds. (eqs-news.com) 11/ The announcement does not disclose pricing, customer names, or rollout dates for the upgraded applications. It is a product-positioning release: Hikvision is trying to show that Guanlan has moved from a model launch into concrete operational workflows tied to alarm quality, searchability and labor reduction. That last point is an inference from the functions the company emphasized in its release and product pages. (eqs-news.com) 12/ The next thing to watch is whether Hikvision publishes customer deployments, benchmarks or product documentation that break out performance by scenario. As of Sept. 8, the public materials around this update are the company announcement and Guanlan product pages describing edge detection, multimodal retrieval and scenario-specific analytics. (eqs-news.com)