The surveillance-video tier list: night-vision IR, perimeter intrusion and cross-camera tracking — the eighth video class nobody benchmarks

Surveillance video: the modality every public benchmark skips

Every public video benchmark tests the same material: film clips, variety shows, YouTube highlights, lecture recordings. Video-MME, MVBench, LongVideoBench — not one tests surveillance footage.

That is not an oversight; it is three structural reasons:

• The data cannot be public. An informative surveillance clip contains faces, plates, premises information and timestamps, constrained by personal-information protection law and security grading — it cannot enter any public dataset.
• Annotation needs security-domain knowledge. The boundaries between “climbing” and “loitering”, “projectile” and “falling object”, “gathering” and “congestion” are defined by security operations standards, not by any academic taxonomy.
• Overseas has no demand. Perimeter defence, cross-camera tracking and safety compliance are Chinese-specific scenarios with no counterpart in the English-language evaluation stack.

Together: surveillance-video understanding has no public score to consult. Yet the scenario has real paying demand — hundreds of millions of camera endpoints nationwide, with campuses, transport, bank branches and industrial sites all buying “models that can read surveillance footage”.

Of the 32 models on the 2026 China native multimodal LLM tier list, exactly one treats surveillance video as its primary training scenario: Hikvision Guanlan 3. This essay builds a separate mini tier list over eight surveillance classes for it — not to rehabilitate it, but to record a skipped modality.

The eight surveillance classes and their difficulties

Class	Core difficulty	Difference from general video
Night-vision IR recognition	IR greyscale, low-light noise, thermal contours	General models’ night priors come from street scenes, not camera positions
Perimeter intrusion	Small targets, long range, frequent occlusion, false-alarm control	Tests “is there” and “is it a false alarm”, not “what happened”
Cross-camera tracking	Re-identification across camera positions, clothing change, lighting jumps	Needs cross-camera identity consistency; single-camera models cannot
Crowd density and counting	Dense occlusion, overhead angles, real-time requirements	General counting priors come from street crowds a magnitude sparser
High-altitude projectiles	Very small targets, fast descent, complex backgrounds	Frame-level localisation plus trajectory recall; second-level timing is mandatory
Fire and smoke	Early small targets, distinguishing sunsets and headlights	False-alarm cost is extreme; precision matters far more than recall
Plate recognition	Motion blur, night glare, grime	A recognition task, not an understanding task — overlapping Congrong’s home ground
Safety compliance	Combined judgement of helmets, workwear, fence crossings	Combined rule judgement; single-frame understanding is insufficient

The third column is the point: five of the eight classes (perimeter, cross-camera, crowd, projectiles, compliance) are not “video understanding” tasks at all but combinations of detection, tracking and rule evaluation. Putting them in the same procurement flow as “understanding video” is the most common mistake in security selection.

Per-class results

Each class lists its top three and main failure modes. Scores are 0–100 desk assessments from vendor and integrator channels, security industry case records, and comparable published metrics. Guanlan is grade C: every figure is vendor-reported with no third-party reproduction — because nobody can reproduce it: there is no public surveillance evaluation set.

Night-vision IR recognition

1. Hikvision Guanlan 3 89 — IR and low light are home ground; no drop-off in low-light scenes
2. Doubao Seed 2.0 Pro 74 — night street priors transfer; IR-specific features missing
3. Qwen3.8-Max 72 — good general night capability; weak on thermal contours

Failure modes: tree shadows, lights and small animals judged as human targets; confidence drift when IR and visible frames are mixed.

Perimeter intrusion

1. Hikvision Guanlan 3 91 — 94.2% security-event recall (vendor figure); false-alarm control is a strength
2. Doubao Seed 2.0 Pro 70 — can describe “someone entered”, cannot grade by security standards
3. Qwen3.8-Max 69 — same, and latency above security real-time requirements

Failure modes: conflating “loitering” with “passing” (two different events in security standards); no security event codes, only natural-language description.

Cross-camera tracking

1. Hikvision Guanlan 3 86 — cross-camera identity consistency is part of the hardware loop
2. Doubao Seed 2.0 Pro 68 — steady within one camera; identity lost across cameras
3. Qwen3.8-Max 67 — same

Failure modes: identity breaks after clothing change; one target judged as two when camera clocks are unsynchronised.

Crowd density and counting

1. Hikvision Guanlan 3 88 — overhead-angle and dense-occlusion counting priors
2. Doubao Seed 2.0 Pro 75 — street-crowd counting transfers; drops in dense bands
3. Step-3 68 — usable at normal density; overcounts in crowds

Failure modes: systematic undercount above four people per square metre; static onlooker crowds misjudged as lingering events.

High-altitude projectiles

1. Hikvision Guanlan 3 84 — small-target descent detection and trajectory recall
2. Doubao Seed 2.0 Pro 72 — frame-level description fine; trajectory continuity poor
3. Keye-VL-2.0 68 — built on video 94; still weak on small targets

Failure modes: raindrops, birds and leaves judged as projectiles; can say “something fell” but not which floor it fell from.

Fire and smoke

1. Hikvision Guanlan 3 87 — early smoke versus flame distinction; false-alarm control first
2. Doubao Seed 2.0 Pro 76 — good on open flame; weak on early smouldering
3. SenseNova 6.5 Pro 71 — professional vision base; lacks security fire corpora

Failure modes: sunsets, headlights and welding sparks trigger false alarms — for fire and smoke the cost of false alarms exceeds misses, so the precision curve matters more than the recall curve.

Plate recognition

1. Cloudwalk Congrong 2 85 — recognition is its old trade; steadiest under motion blur and night glare
2. Hikvision Guanlan 3 83 — completed inside the hardware loop; slightly behind Congrong on pure model figures
3. Doubao Seed 2.0 Pro 78 — usable on clear plates; drops on grime and blur

Failure modes: character completion on damaged plates (the model will “guess” a plausible plate — the most dangerous failure mode in recognition); full overexposure under direct headlights.

Safety compliance

1. Huawei Pangu 5.5 80 — most industrial deliveries; the most complete rule library
2. Hikvision Guanlan 3 78 — steady on helmets and workwear; average on combined fence-crossing judgement
3. SenseNova 6.5 Pro 76 — strong visual detection; weak rule combination

Failure modes: sun sleeves and headscarves judged as violations; reports “someone without a helmet” but not location or duration — yet those are the fields a compliance ticket needs.

The mini board: almost unrelated to the general video ranking

Weighting the eight classes by security business frequency (perimeter 20%, crowd 15%, IR 15%, fire 12%, compliance 12%, cross-camera 10%, projectiles 8%, plates 8%):

Rank	Model	Weighted	Overall	Video score	Gap
1	Hikvision Guanlan 3	86.5	#27	91 (7th)	−26
2	Doubao Seed 2.0 Pro	72.9	#2	96 (1st)	−1
3	Qwen3.8-Max	71.3	#1	93 (4th)	+2
4	SenseNova 6.5 Pro	68.7	#8	89	+4
5	Huawei Pangu 5.5	68.6	#21	80	+16
6	Cloudwalk Congrong 2	67.4	#32	68 (no video modality)	−26

Three conclusions:

• 27th overall is first here; 1st overall is third. General video scores have almost no predictive power for security procurement — Guanlan’s video 91 versus Doubao’s 96 is a 5-point gap, but 13.6 points on this mini board.
• There is exactly one professional. This is the first board on this site with only one professional model — the other five are general models doing security on the side. They can “describe” surveillance footage; they cannot emit security event codes, false-alarm grades or ticket fields.
• The model with the lowest video score (Congrong at 68 — in fact no video modality) ranks 6th. Its plate recognition 85 is second in class — a model with no video modality reaching sixth on a surveillance board through recognition alone. Once again: “multimodal” is too coarse a word.

The product form of surveillance video: minute-level recall, not second-level timing

Guanlan’s TTA (temporal tolerance area) is only 54.0, 9th of 32 in long video understanding — flat on coarse events, steep at second precision. 92% at ±60s, only 18% at ±1s.

That is not a defect; it is the product form. Security review needs “the climb happened at what minute, lasted how long” — minute level suffices; second-level timing is for sports editing and forensics, a bonus in surveillance. Judging a summary product by timing-product standards (Doubao’s 68% at ±1s) is like grading a calendar with a stopwatch.

The procurement implication is direct: if your business is surveillance, the acceptance standard should be “minute-level event recall + false-alarm rate”, not any general video benchmark score. On the former Guanlan is 94.2% (vendor figure); on the latter it is 9th. Both numbers are right; they answer different questions.

The acceptance checklist: 200 clips from your own sites

Trust no vendor demo reel. Pick 200 clips from your own sites in this ratio:

• 80 typical samples (half day, half night, covering all camera types)
• 50 boundary samples (IR mode, backlight, rain and fog, low light, lens grime)
• 40 adversarial samples (small animals, tree shadows, lights, drones, loitering people)
• 30 “should refuse” samples (too blurred for a human, target fully occluded)

The last group is the most overlooked. In security scenarios the cost of a false alarm exceeds a miss, and a confident misjudgement is an order of magnitude more dangerous than a refusal. Record three metrics per class: event recall, false-alarm rate (per camera per day), and output-field completeness (can it emit event code, location, duration). An acceptance that reports only the first has no reference value.

Compliance gates: security grading, personal-information law and data residency

Security gates have nothing to do with capability but matter more:

• Data residency. Surveillance footage contains personal information and in principle never leaves the customer’s premises — which excludes all pure cloud API solutions outright; Guanlan’s industry licensing (edge box + central cloud) is designed for exactly this.
• Separate compliance for faces and plates. Two categories of biometric information have separate collection and use rules; if model output includes structured face features or plate-recognition results, a corresponding compliance path is required.
• Security grading and logs. Security systems typically run under graded protection requirements, and model calls and outputs must be auditable — which requires vendor log interfaces, not black-box APIs.

The tier list tells you which models can read surveillance footage, not which one may enter your machine room. The former is this essay’s mini board; the latter is G2 of the veto-gate tier list — where Guanlan precisely dies (industry licensing; the client never holds weights). Its correct procurement identity is “an industry-licensed product delivered with hardware and solutions”, not “a self-held model asset”. Those two must be written separately in the tender.

Boundaries and method

The eight-class scores are desk assessments from vendor and integrator channels (grade C), with no third-party reproduction, because no public surveillance evaluation set exists. The weighting is set by us according to security business frequency and is disputable. Guanlan’s dossier and per-item evidence are on the model page; its general dimension scores and TTA curve are in long video understanding.