- New multimodal model combines vision and language reasoning for road monitoring
- Company targets police and highway operators amid push for smarter infrastructure
Chinese surveillance and AI juggernaut Hikvision (海康威视) has launched a multimodal traffic incident detection system that uses large AI models to interpret road scenes rather than simply identify visual anomalies, aiming to reduce false alarms in highway monitoring.
The system combines computer vision and language-based reasoning to detect traffic incidents including debris, illegally parked vehicles and pedestrians entering highways, according to the company.
Traditional traffic monitoring systems often generate high false-positive rates because they rely on pixel-level image changes without understanding contextual relationships in a scene.
Reflections from road signs, for example, can be mistaken for stopped vehicles, while water stains on pavement may trigger debris alerts.
Hikvision said its new system addresses those problems through a two-layer architecture. A vision model first identifies objects such as vehicles, pedestrians and obstacles, while a language model analyzes the logical relationships between them.
The platform also uses a two-stage verification process. A visual model conducts high-sensitivity preliminary screening to identify potential incidents, after which a multimodal AI model performs deeper analysis using contextual information including location, environment and timing.
90% of false alarms
The company said the process filters out more than 90% of false alarms before alerts are sent to human operators.
For example, if the system detects a tire on a highway lane, it can determine that the tire is detached from a vehicle and lies on the roadway rather than mounted on a moving car, allowing it to classify the object as road debris instead of a normal traffic condition.
Hikvision said the product is based on its Guanlan large-model framework introduced in August 2025 and represents a key deployment of the company’s multimodal AI strategy in transportation infrastructure.
The company has also developed road-specific training models to improve detection accuracy for scenarios that have historically been difficult for highway systems, including scattered cargo, illegal parking and pedestrian intrusions.
Target users include public security traffic police, highway patrol authorities and expressway operators across China.
The system also supports few-shot learning, enabling it to generate accurate alerts without requiring massive labeled datasets.
Hikvision is offering both edge-based and centralized deployment options. At the edge, cameras integrate dedicated AI image-processing algorithms designed for road environments, including suppression of nighttime headlight glare, while performing real-time local analysis.
For centralized deployments, incident detection servers can integrate with existing video monitoring infrastructure, allowing operators to reuse legacy systems and reduce upgrade costs.
Covering 130,000 kilometers
Hikvision said its broader digital highway infrastructure business now covers more than 60 application scenarios spanning toll stations, tunnels, bridges, slopes and service areas.
The company’s transportation solutions have been deployed across more than 130,000 kilometers of highways nationwide.
