False Alarms Are the Real Battleground: The Evaluation Revolution Hidden Behind Scene Understanding

Everyone at the show talked about models that understand what they see. What actually decides whether AI scales in fleets is how often it alerts when it should not. A municipal transport authority has already put that phrase into an official reply.

The dominant narrative at Security China 2026 was cognitive intelligence — systems that understand scenes rather than detect objects. But the most commercially useful sentence of the whole show came from a closing report on infrared thermal imaging: a high false alarm rate is the core obstacle to scaling thermal imaging in security. For fleet video, that sentence is worth more than any demo.

The Evidence on the Floor

Infrared specialist InfiRay stated the problem plainly and offered a two-layer answer: front-end initial screening plus back-end precise adjudication, with conventional algorithms and large-model reasoning running in parallel. In forest-fire scenarios the system filters sunlight reflection, greenhouses and glass or metal glare; in warehouses it distinguishes genuine abnormal temperature rise in stacks from transient heat sources such as forklift engines, stacker cranes and passing personnel.

The hardware claims are specific: 8-micron and 12-micron vanadium oxide detectors, thermal sensitivity better than 15mK, 640x512 infrared resolution, 256x192 detectors with 4x super-resolution, a turntable accurate to plus or minus 0.01 degrees, 100 degrees per second slew, and a claimed 95 percent first-attempt success rate when radar cues the pan-tilt unit. Those numbers describe a direction, not a benchmark — they are manufacturer claims, not independent tests.

The wider industry is converging on the same problem. Pre-show technical surveys cite an 11-TOPS dual-lens AI bullet camera with an all-weather recognition rate above 90 percent, a training-and-inference appliance that cut algorithm deployment from weeks to a single day, and a large-model security protection system from a Ministry of Public Security research institute with over 90 percent average detection accuracy. The State Council AI+ action opinion explicitly calls for a model capability evaluation system.

Regulators Measure Closure, Not Accuracy

Here is the mismatch that matters. Vendors quote 99 percent recognition accuracy. The regulatory system for commercial vehicles measures three other things: timely handling, dynamic clearance of outstanding alerts, and a 90 percent handling rate. Once false alarms proliferate, front-line monitors drown in noise and genuine high-risk events get missed. The operator either hires more people or turns down sensitivity — the first raises cost, the second collides with rules that forbid altering alert thresholds without authorisation.

The problem is documented. In an April 2026 reply, the Dongying municipal transport authority named the relatively high false alarm rate of on-board intelligent video terminals as a widespread issue in hazardous-goods road transport and committed to working with technology companies to upgrade and optimise those terminals. The four-ministry action plan requires timely handling and dynamic clearance of alerts, penalises enterprises below 90 percent handling, and prohibits both unauthorised changes to alert limits and falsification of monitoring data.

In a fleet context, a false alarm is therefore not a user-experience issue. It is a compliance issue. A system producing 30 invalid alerts per thousand kilometres drives up labour cost, handling cost and regulatory exposure at once, however flattering its single-event recognition metric looks.

Our Take

The front-end screening plus back-end adjudication pattern transfers almost literally to vehicle architecture. A lightweight model on the unit performs first judgement and filtering; only structured summaries of suspect events — not full video — go upstream; a cloud or regional node uses a larger model for second judgement and attribution; confirmed false alerts flow back into the training set and update on-device thresholds through OTA. That caps bandwidth cost and turns false-alarm reduction into a continuously converging engineering metric instead of a one-off tuning exercise.

Just as important, an alert must carry its own evidence: video clip, time and location tag, triggering rule, model version, and driver and vehicle context. Without those, a handler cannot judge whether it is real, and the closed loop stops at read.

Seven Metrics Worth Putting in a Tender

• Valid alerts per thousand kilometres — denominator is real mileage, aligned directly with fleet operating cost and harder to game than recognition accuracy.

• False alarm and miss rates by scenario — report separately for night, backlight, rain and fog, tunnels and rough roads; forbid a single global average.

• Attributable alert ratio — share of alerts carrying a clip, spatiotemporal tag, rule and model version; this determines whether closure is possible.

• On-device first-judgement latency — fatigue, distraction and forward-collision interventions must be decided locally, not after a round trip.

• Offline availability — reproducible tests for local judgement, buffering, resumable upload and degraded behaviour with no network.

• Model OTA and rollback — whether updates are grey-released, reversible and version-traceable, which directly sets the speed of false-alarm convergence.

• False-alarm convergence curve — require the supplier to commit to reduction targets at one, three and six months, turning continuous optimisation into a contractual obligation.

Our Judgment

The 2027 dividing line in fleet AI will not be whether a product has a large model. It will be whether false alarms can be reduced into a curve that is measurable, attributable and iteratively improvable. Whoever rewrites the evaluation language from model accuracy to valid alerts per thousand kilometres plus attributable alert ratio will redefine the rules in regulated procurement. That is precisely what a national model capability evaluation system should mean at industry level.

Sources: China Economic News Network closing report on infrared plus AI, pre-show technical surveys, the State Council AI+ action opinion, Jiaoyunfa 2026 No. 2 and local implementation documents, and the Dongying municipal transport authority reply of 13 April 2026. Vendor parameters are public claims without third-party verification or same-condition comparison and should not be used directly as tender scoring data.