Relying onHikvision's Viewlan Large Model Technology System

HikvisionDeploy large model capabilities directly to the edge

Introduce a series of visual large model cameras

The visual large model camera not only produces superior image quality.

What's more important

Overcame the weak generalizability of scenarios.

Difficulties in complex target recognition and other bottlenecks

Enhance Target Detection Accuracy Further

Significantly Reduce False Positives

The system can also recognize in zero-shot scenarios.

A wider variety of detection targets

Significantly Reduce False Positives

Detection rates and accuracy rates both improve

Smart cameras, in practical application, can lead to numerous false alarms due to varying scenarios, lighting conditions, and weather, which increase operation and maintenance costs, degrade user experience, and diminish trust in event responses.

For over two decades, Hikvision has been deeply rooted in the video domain, leveraging a wealth of industry knowledge across various scenarios to build a pre-trained large model. In the pre-training phase, we've augmented the data with real dynamic scene disturbances such as rain, snow, fog, bright flashes, animal movement, and vibrations, enhancing data under different conditions to significantly improve the detection and accuracy rates of intelligent recognition.

Hikvision has also established a comprehensive large model deployment technology system. It has researched model structure design and quantization technology from aspects such as model lightweighting, improved computational efficiency, and saved computational resources, innovatively developing visual large model cameras that are more adaptable to specific scenarios.

In perimeter applications, Hikvision has developed a large model alert series of cameras, including PTZ cameras, IPCs, and multi-sensor models. Compared to traditional perimeter video products, the large model alert series cameras have significantly improved the recognition distance and reduced the false alarm rate by over 90%.。(Based on actual project test data)

Under the same test scenario, a 4mm lens was used for testing.The visual large model camera can detect human intrusions at a distance of 70 meters.Historically, deep learning algorithms could detect at 40 meters, while traditional smart algorithms could only detect at 20 meters.

* In the same scenario, using a 4mm lens for testing, the traditional smart algorithm detected human intrusion while continuously detecting bird intrusion.The visual large model camera accurately filters false alarms from birds, detecting human intrusions only.

In traffic event detection, Hikvision has launched visual large model cameras including the RayVision all-in-one machine, event detection cameras, and the FOD RayVision all-in-one detection machine. In the field of highway traffic event detection, these cameras effectively address the challenges of false and missed detections for incidents such as spilled materials, parking, and pedestrians in complex scenarios.

In traffic checkpoint applications, we have launched visual large model camera products such as checkpoint capture units, non-motorized vehicle capture units, and the Leiyun ship checkpoint integrated machine. For instance, in the cabin feature recognition application, it effectively filters false alarms caused by low contrast, obstructions, and complex postures when identifying seat belts; and it also filters false alarms caused by raising hands or holding objects when identifying phone calls.

Support for Zero-Sample Open Recognition

A wider range of target identification types

The implementation of traditional intelligent applications faces diverse intelligent needs across various industries. Tailoring specific recognition algorithms to different targets involves challenges such as high sample collection costs, difficult category expansion, and a lengthy training cycle.

Hikvision has launched a smart camera application mode named "Description to Recognition." By deploying an open-source target detection large model on cameras and adopting a self-developed unified modality learning solution, it aligns image features of visual recognition with semantic features, achieving precise detection and localization of targets. This innovative mode allows for the quick and flexible generation of models with just a word or a sentence input, identifying targets without the need for sample training.

Scene Definition Image Quality

Precise Adaptive Optimization

The visual large model camera integrates "hardware + algorithm" in a deep fusion to enhance image quality across all scenarios. With professional large-aperture lenses, high-sensitivity sensor design capabilities, and extensive experience in low-light environments, it constructs an end-to-end intelligent large model algorithm that effectively distinguishes signals from noise in images, achieving precise noise reduction. This boosts the signal-to-noise ratio in night vision surveillance, and for scenarios like heavy rain, smog, overexposure, and color cast, it allows for scene-defined image quality, resulting in richer image details and more authentic colors, providing superior video image support for intelligent applications.

The advent of large models has further enhanced Hangzhou Hikvision Digital Technology Co., Ltd.'s technological and product innovation capabilities. We will continue to delve into various scenarios to address practical issues for our customers, facilitating the intelligent upgrade of applications across numerous industries.

Currently, Hikvision has launched

Alert Series, Traffic Incident Detection Series,

Knot Capture SeriesSeries of anti-collision mechanical arms

Vibration and Shock Resistance Series, Inspection SeriesPlease provide the Chinese content that needs to be translated into American English.

Visual Large Model Camera