Chinese firm Hikvision is combining AI and IoT in a suite of checkpoint cameras, traffic incident detection cameras and servers.

These are powered by large-scale traffic AI models and are designed to enhance detection accuracy, reduce false alarms and improve overall traffic safety.

Hikvision’s large-scale traffic AI models are based on the Guanlan large-scale AI model (www.hikvision.com/en/core-technologies). They have been developed with deep traffic industry knowledge and are trained on millions of complex real-world situations such as rainy roads, construction areas and extreme weather.

The models use the Transformer architecture, an AI-building approach that allows simultaneous viewing of data parts. It assists the models by identifying complex features and can rapidly determine the relationships between different elements in a traffic scene.

The self-attention mechanism in these models enables better extraction of overall features. Instead of focusing on individual details, they analyse the whole picture to understand the context. This reduces false detections caused by the background.

The checkpoint cameras with large-scale AI models (www.hikvision.com/en/products/ITS-Products/Checkpoint-Systems) address common traffic compliance monitoring problems such as detecting seatbelt use and phone usage by drivers.

For seatbelt detection, the self-attention mechanism evaluates the driver’s posture and visible parts of the seatbelt, even when partially obscured, and reduces false alarms by 75%. The models’ strong pre-training generalisation and self-attention mechanisms allows them to work well in complex situations, such as when the seatbelt colour matches the clothing or the belt has an unusual shape.

When it comes to detecting phone use, these cameras analyse multiple factors such as hand position, the area near the ear and the driver’s line of sight. By integrating industry knowledge rather than relying solely on the phone’s shape, they are said to reduce false phone use detections by 75%.

The incident detection cameras and servers (www.hikvision.com/en/products/ITS-Products/AID-Systems) leverage large-scale AI models to deliver a 60% reduction in false alarm rates for incidents such as fallen objects, pedestrian detection and stopped vehicles.

The cameras reduce background false detection by better understanding the image context. Their enhanced generalisation ability allows them to detect rare targets and capture complex features, such as identifying event types based on key characteristics of personnel and vehicles. For example, they can identify reflective vests or emergency lights to classify events more accurately. The servers, in turn, let users get these benefits using their existing cameras, avoiding the need to install new ones for road monitoring.

“These new cameras are a big step forward in traffic management efficiency,” said Hollis Li, intelligent traffic product director at Hikvision (www.hikvision.com). “With large-scale traffic AI models, our offerings are more accurate and better at handling real-world challenges. This means safer roads and more efficient traffic management for everyone.”