Camera Intelligence — October 2026: Crime, Fights & Assaults

Five surveillance recordings. Two plain-language Events. How four models respond to fights and assaults, and when they stay quiet.

Recognising the moment a confrontation becomes physical

Two short Events: “People are fighting with each other.” and “A person physically attacks another person.” We tested them on five selected surveillance recordings: three fights and two assaults. The question was whether each model noticed the requested event while the incident was taking place.

All four models received identical frames. A later match still counts as detection. These are selected examples from UCF-Crime, not a general crime-detection accuracy score or a judgement about the people shown.

A confrontation in a busy entrance

Event: “People are fighting with each other.” People pass through the scene before physical contact begins. The reference incident starts around one minute into the recording.

• Footage source: UCF-Crime

A brief fight in a passageway

Event: “People are fighting with each other.” A brief confrontation is followed by a person remaining on the ground. TypeSafe, Liquid and Qwen also mistook the closing martial-arts illustration for a fight. That extra false alert is not included in the table’s scored footage.

• Footage source: UCF-Crime

A fight further down a corridor

Event: “People are fighting with each other.” Several people are visible, but the physical interaction occupies only part of the view.

• Footage source: UCF-Crime

An assault around a counter

Event: “A person physically attacks another person.” Physical contact is intermittent in this crowded scene. Qwen also raised an alert during ordinary activity before the attack.

• Footage source: UCF-Crime

An attack on a street

Event: “A person physically attacks another person.” The scene includes an attack and people moving through the aftermath.

• Footage source: UCF-Crime

Quiet scenes and later matches

The same normal-activity excerpt was checked against both Events. OpenAI gpt-6-luna: 0% false alerts; TypeSafe AI Jev 1.14.0: 0% false alerts; LiquidAI d1-2B (bfloat16): 0% false alerts; Qwen3.5-9B (Q4_K_M): 0% false alerts. This is a narrow control, not a field false-alarm rate.

An alert after an attack can reflect continuing activity or a mistake. In the passageway example, TypeSafe, Liquid and Qwen also alerted on a closing illustration rather than a real fight.

What these results can tell us

¹ Detected means at least one match during the incident. Unknown means the model was uncertain. These selected examples do not establish general accuracy.

² False alerts count alerts during the scored scenes without an incident. Aftermath and closing graphics are excluded from these percentages; the extra alerts on graphics are described above.

³ Memory figures show peak usage on an RTX 4080 Laptop GPU, including model loading. Liquid and Qwen use different GPU memory measurements, so these figures are approximate comparisons, not minimum hardware requirements.

Response times measure the model response, not the time from an incident to an alert. Cloud costs are estimates; local hardware and electricity are excluded. “—” means unavailable or not applicable.

• Download all observations and measurements (JSON)