VLM output
An elevated highway camera looks down on a dry, multi-lane carriageway in daylight where several dark pieces of debris lie scattered across the rightmost lane near the edge line. Cars and a dump truck continue past at speed, some passing directly over or beside the debris without stopping, and one dark object is thrown up alongside a white car in the first frame.
- Road Hazard PresentYes
- Hazard TypeRoad Debris
- Debris Pieces Visible5-8
- Unsafe Behavior ObservedYes
- Unsafe Behavior TypeDebris Straddling
- Pedestrians On RoadwayNo
- Collision OccurredNo
- Vehicles In View8-12
- Traffic FlowFree-Flowing
- Lighting And WeatherDaylight, Dry
Customer challenges
Only crashes are recorded; near misses and dangerous behaviour in the footage are never captured
Legacy CV cannot read the interaction of people, vehicles and road context as risk
Failure-to-yield, box blocking and obscured signals are not detected systematically
Managers know where crashes cluster but have no visual evidence of why
Road improvements cannot be verified before and after with observed behaviour
Safety education lacks realistic scenes reconstructed from real intersections
Customer benefits
From post-crash outcome analysis to pre-crash risk analysis
A Near Miss Dataset that did not exist before
Identifies the truly high-risk locations and behaviour patterns, with video evidence for road improvements
Quantifies the change in dangerous events before and after improvements, with AI continuously validating results
Lower cost of watching large CCTV estates by hand
Performance defined by crash prevention (A1 and A30 reductions), establishing data-driven road safety management
Digital-twin education material lets road users see the risk before the crash
Key features
Failure-to-Yield to Pedestrians Detection
Detects vehicles failing to yield to pedestrians.
Intersection Accident Analysis (Type, Vehicles, Injuries, Responders)
Analyzes intersection accidents, including type, vehicles involved, injuries, and responders.
Near-Miss Detection (Vehicle–Pedestrian Conflict, Hard Braking)
Detects near-misses such as vehicle–pedestrian conflicts and hard braking.
Dangerous Driving Behavior Detection (Wrong-Way, Illegal Parking, Pedestrian Intrusion)
Detects dangerous driving behavior including wrong-way driving, illegal parking, and pedestrian intrusion.
High-Risk Intersection Condition Detection (Box Blocking, Roadwork Narrowing, Obscured Signals)
Detects high-risk intersection conditions such as box blocking, roadwork narrowing, and obscured signals.
Risk Hotspot Mapping & Safety Insight Generation (Digital Twin)
Maps risk hotspots and generates safety insights in the digital twin.
Road Safety Education Video Generation (Digital Twin)
Generates road safety education videos from digital twin scenarios.
VLM Safety Understanding
Uses a vision language model to interpret safety-relevant scenes.
Dangerous Behavior Detection
Detects dangerous behavior in the monitored area.
Incident Understanding
Interprets what happened in a recorded incident.
Risk Event Classification
Classifies risk events by type.
Safety Digital Twin
Provides a safety digital twin of the monitored site.
AI Safety Agent
Provides an AI safety agent to assist review and response.
Safety Insight Generation
Generates safety insights from accumulated events.
Before & after
| Before | After — with Linker AI Nexa |
|---|---|
| Before 0% of near misses recorded | After — with Linker AI Nexa Continuous near-miss capture on 100% of monitored intersections |
| Before Crash awareness via citizen calls | After — with Linker AI Nexa 80% proactive detection, responder arrival -50% (crash scenario) |
| Before CCTV reviewed only after a crash | After — with Linker AI Nexa 60–125x more feeds under continuous analysis |
| Before Failure-to-yield evidenced by occasional patrols | After — with Linker AI Nexa 24/7 evidence on 100% of monitored crossings, with weather, severity and pedestrian type |
| Before Improvements judged by crash counts years later | After — with Linker AI Nexa Before/after change in dangerous events measured within months |
| Before Safety education built on generic material | After — with Linker AI Nexa Scenes reconstructed from real intersections, already used in licensing and campus courses |