Hikvision's groundbreaking large vision model has recently made its mark in the transportation sector. Leveraging the Guanlan Large Model to transcend traditional algorithmic constraints, Hikvision has unveiled a new generation of edge-based event detection cameras and simultaneously integrated large model capabilities at the central hub, introducing event detection terminals and servers. This holistic approach, spanning from the edge to the center, significantly enhances the performance and efficiency of intelligent applications, accelerating the intelligent transformation of the transportation industry.

In highway traffic event detection, industry bottlenecks have persisted due to limited model scale and generalization ability, leading to false and missed alarms for events such as scattered objects, parking violations, and pedestrian intrusions in complex scenarios.

Drawing on years of industry expertise and leveraging the Guanlan Large Model's technical foundation, Hikvision has developed data models tailored to industry-specific scenarios like road events. These models are deeply integrated with intelligent hardware, resulting in a new generation of event detection products that accurately detect road anomalies, thereby ensuring road safety and smooth traffic flow.

Enhancing Road Intelligence through Comprehensive Understanding

Traditional traffic event detection algorithms heavily rely on manually labeled scenario data, necessitating separate sample collection and model training for each event type. This often leads to missed detections and misjudgments due to incomplete sample coverage in real-world applications. For instance, ponding on roads during rainy weather might be misidentified as scattered objects.

Consider a highway segment equipped with 1,500 cameras. Traditional algorithms can detect over 1,000 events per day on average. The key to improving highway operational efficiency lies in reducing invalid events, minimizing manual review workloads, and accelerating the response time from detecting to handling genuine abnormal events.

Utilizing the Guanlan Large Model, Hikvision employs industry knowledge pre-training and fine-tuning to equip the model with expert-level capabilities in event detection. Compared to traditional convolutional networks, the Transformer-based model boasts a deeper architecture, stronger global feature extraction, and superior context modeling abilities, resulting in enhanced model generalization and systematic resolution of false and missed alarms in complex scenarios.

For instance, in scattered object detection, the large model significantly improves detection accuracy, mitigating interference from tree shadows, water stains, markings, and signboards. In parking detection, it precisely distinguishes signboards, slow-moving vehicles, and construction vehicles, comprehensively assessing dynamic features like vehicle dwell time and lane deviation, drastically reducing false alarms.

Comprehensive Product Matrix for Multi-Scenario Needs

Powered by the large model, Hikvision's event detection series products enhance the comprehensive understanding of the physical world by extracting valuable information from multi-dimensional signals and mining potential relationships between different modal data. This breakthrough transcends performance limitations in diverse environments, including day and night, rain, and fog, facilitating large-scale application deployments across edge and central systems.

I. Edge-side Sensing Devices: Second-level Detection, Precise Identification

The event detection radar-camera integrated machine leverages millimeter-wave radar's all-weather high-precision ranging and speed measurement capabilities to achieve second-level preliminary detection of illegally parked vehicles and road anomalies in extreme conditions like low light, rain, and fog. Subsequently, through large model reasoning, it performs decision-level fusion of multiple feature types, satisfying the need for high detection rates and accuracy in complex environments.

The event detection camera is equipped with a road-specific AI ISP image processing algorithm that effectively mitigates the impact of vehicle high beams on images, ensuring night images remain bright and clear. Coupled with the large model, it conducts fine-grained classification and filtering of events like scattered objects and pedestrian intrusions, performing real-time analysis locally and significantly reducing false alarm rates. Relying on device-side computing power, it eliminates cloud dependency and substantially reduces network transmission pressure.

II. Center-side Equipment: Reusing Old Resources for Cost Reduction and Efficiency Enhancement

Event detection terminals/servers enhanced with large model reasoning capabilities seamlessly integrate with existing video sensing systems and storage devices, enabling the reuse of existing resources. Edge nodes focus on real-time event detection, while central servers handle multi-source data correlation, ensuring low-latency response while meeting the intelligent needs of varying-sized scenarios.

Based on Hikvision's Guanlan Large Model technology system, the traffic event detection series products elevate the performance and effectiveness of intelligent applications. Looking ahead, we will continue to explore large model technology, integrating the Guanlan multimodal large model's image-text understanding and reasoning capabilities to expand application scenarios like road accidents, collapses, and abnormal weather conditions. This will further our understanding of roads and continuously drive the intelligent evolution of the industry.