Camera Intelligence — October 2026: Road Accidents & Everyday Traffic

Three collisions, two normal street scenes, and one plain-language Event. A comparison of what four models detect—and when they stay quiet.

Not every moving vehicle is an accident

One Event: “A road traffic accident occurs.” We tested it on three collisions and two normal street recordings, using the same frames for every model.

A collision only needs one match during the accident to count as detected. Normal clips are scored separately: staying quiet matters too. These five selected examples are a small case study, not a general accuracy estimate.

A distant collision at a junction

The collision happens quickly and occupies a small part of a distant, low-resolution view. Its exact moment may be hard to see.

• Footage source: UCF-Crime

A collision beside a pickup

The accident happens beside a pickup near a crossing. Liquid and Qwen also alerted during the aftermath; those later alerts are not counted as detecting the collision itself.

• Footage source: UCF-Crime

A bus hits a parked car

The bus enters the view and pushes a stationary car out of position. The change is visible across several consecutive images.

• Footage source: UCF-Crime

Daytime traffic without a collision

Vehicles move along the road without a visible collision. This scene checks whether ordinary traffic produces an alert.

• Footage source: UCF-Crime

A nighttime street without a visible crash

People in reflective clothing stand or walk near stationary vehicles. No collision is visible. Their appearance alone is not evidence of a road accident.

• Footage source: UCF-Crime

What changes from one scene to the next

Across the three accident events: OpenAI gpt-6-luna: 33% detected; TypeSafe AI Jev 1.14.0: 67% detected; LiquidAI d1-2B (bfloat16): 0% detected; Qwen3.5-9B (Q4_K_M): 67% detected. These percentages describe only the three selected collisions, not performance across road traffic in general.

Both normal recordings are kept separate from event detection. OpenAI gpt-6-luna: 0% false alerts on scenes without an accident; TypeSafe AI Jev 1.14.0: 0% false alerts on scenes without an accident; LiquidAI d1-2B (bfloat16): 0% false alerts on scenes without an accident; Qwen3.5-9B (Q4_K_M): 50% false alerts on scenes without an accident. Two short controls cannot establish a real-world false-alarm rate. A zero false-alert rate can also include unknown responses; the per-case outcome shows these explicitly.

Distance and sampling matter. Every model received the same low-resolution images at one-second intervals. A missed event may reflect the model, the visual evidence available to it, or both. This experiment does not separate those causes.

Reading the results

¹ Detected means at least one match during the collision. Later alerts during the aftermath do not count as detecting the accident itself. Unknown means uncertainty, not a correct answer.

² False-alert percentages describe the two scenes without a collision. This small set cannot establish a general false-alarm rate.

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)