The new SmartVision 6.4 release takes another step away from conventional video surveillance and toward a full-scale computer vision platform. The update introduces Advanced Object Detection, support for multiple YOLO models, and optimized parallel processing of video streams from large numbers of IP cameras.
Modern video surveillance is no longer limited to recording what happens in front of a camera. A system must understand what appears in the frame, distinguish a person from a vehicle, detect faces, and recognize license plates. However, the more accurately artificial intelligence analyzes a scene, the more computing resources it requires.
SmartVision 6.4 allows users to choose their own balance between processing speed, detection accuracy, and system load.
Advanced Object Detection
SmartVision 6.4 introduces an additional Advanced Object Detection mode designed to improve object detection accuracy in video streams.
The new mode is primarily intended for computers equipped with a GPU. More sophisticated neural network models can analyze images more accurately, but they also require more video memory and processing power. No miracles have happened yet: the more carefully artificial intelligence studies every frame, the more expensive that frame becomes in computational terms.
The yolomodel.txt configuration file now allows users to select one of four YOLO models. These models differ in weight size, detection accuracy, and hardware requirements.
Lightweight models provide faster processing and are better suited to systems with large numbers of cameras or less powerful GPUs. Heavier models require more resources, but they can potentially detect small, distant, or partially obscured objects more accurately.
The addition of new models and neural network weights has increased the size of the SmartVision installation package. This is the unavoidable price of being able to choose an algorithm for a specific hardware configuration instead of using a single universal model for every deployment.
Parallel Recognition on the GPU
One of the key changes in SmartVision 6.4 is the optimization of the recognition module for multithreaded GPU video processing.
When a surveillance system works with several dozen IP cameras, the load comes from more than video decoding. The system may simultaneously perform object detection, face detection, biometric matching, and vehicle license plate recognition.
SmartVision settings now allow users to specify the number of parallel threads used for face and license plate recognition. This makes it possible to distribute the workload more efficiently across multiple cameras and make better use of modern multicore processors and GPUs.
On powerful graphics cards, the number of threads can be increased to accelerate processing across multiple video channels. On less powerful systems, the value can be reduced to maintain stability and keep resource consumption predictable.
However, the maximum number of threads does not always produce maximum performance. Too many parallel tasks can exhaust GPU memory, create processing queues, and reduce overall throughput. The optimal setting depends on camera resolution, frame rate, the selected recognition models, and the amount of available GPU memory.
ONVIF or Universal H.264
When using resource-intensive recognition models, it is recommended to select reader = onvif in the settings.
In this mode, SmartVision receives the original video stream directly from the IP camera and does not consume additional resources by transcoding it. In systems with large numbers of cameras, this can significantly reduce CPU load and leave more computing capacity available for neural network processing.
There is, however, a trade-off. If a camera transmits video using H.265, the resulting recordings may not open correctly in some web browsers. This is not a SmartVision limitation. It is caused by limited H.265 support and licensing policies implemented by browser and operating system vendors.
When reader = default is selected, SmartVision saves video in H.264. This remains the most universal format and offers broad compatibility with browsers, media players, and mobile devices.
The downside is that the original video stream must be decoded and then encoded again. With only a few cameras, the additional load may be almost unnoticeable. In a large surveillance system with dozens of active channels, however, transcoding becomes a separate computational workload that may compete with AI analytics for system resources.
The appropriate option therefore depends on the deployment scenario.
reader = onvif is better suited to high-performance systems with large numbers of cameras, especially when computing resources need to be reserved for object, face, and license plate recognition.
reader = default is recommended when maximum compatibility of the video archive with browsers and third-party applications is the main priority.
Video Surveillance Becomes a Computing Platform
SmartVision 6.4 demonstrates how the architecture of modern security systems is changing. In the past, the main characteristics of a video surveillance system were camera resolution, storage capacity, and archive retention time. Today, the size of a neural network model, available video memory, the number of recognition threads, and GPU throughput have become equally important.
Cameras continue to deliver images, but an increasing share of the work is performed by software. The software decides which object is important, which event should be recorded, and which frame deserves the operator’s attention.
The new SmartVision release provides flexible control over this computational pipeline. Users can choose between lightweight and heavyweight YOLO models, configure the number of parallel recognition threads, and decide what matters most for a particular system: minimizing hardware load or maximizing video archive compatibility.
SmartVision 6.4 turns these parameters from hidden technical details into practical tools for managing AI-powered video surveillance. A modern security system should not simply watch what is happening. It should understand what it sees and do so simultaneously across every connected camera.