In today’s world, video surveillance systems are an integral part of security both at the level of government institutions and in the private sector. With the advancement of technology and increasing demands on the quality and functionality of security systems, there is a need to integrate more sophisticated and smarter solutions. Artificial intelligence (AI) and machine learning have provided new opportunities to enhance video surveillance systems, allowing them to not only record video but also analyze the resulting data in real time. The origins of video surveillance systems can be traced back to the early 20th century, when the first mechanical surveillance cameras were invented. These early systems were mainly used to monitor public places such as streets and parks. In the 1940s and 1950s, electronic surveillance systems were developed that allowed video signals to be transmitted over long distances. The 1960s and 1970s saw the advent of digital video recording, which revolutionized the video surveillance industry. Digital video surveillance systems provided higher image quality, more storage capacity, and the ability to access video data remotely.

The first applications of artificial intelligence in video surveillance systems centered around recognizing static objects and simple motion patterns. These early systems could alert on unattended luggage left unattended at airports or the detection of motion in restricted areas. The development of these algorithms began with simple image processing techniques such as contour extraction, segmentation, and classification of objects based on their shape and size. However, these methods were limited in their effectiveness because they required clear separation of foreground and background objects, which is not always possible in complex and dynamic scenes. Thus, the progress of machine learning technologies, in particular deep learning methods, has dramatically changed the approach to analyzing video data. Neural networks became capable of training on large amounts of data, which allowed them to recognize not only simple patterns, but also complex behavioral scenarios, human facial expressions and gestures, even in conditions of poor visibility or in the presence of other interference [1]. With the advent of convolutional neural networks, the situation has changed dramatically. These networks are capable of extracting features at different levels of abstraction, which allows them to effectively recognize objects and scenes even in the presence of partial overlap, lighting changes, and other visual distortions. These properties have made convolutional neural networks a key tool in modern machine vision systems.

Figure 1. Device of convolutional neural networks

Converged neural networks analyze pixels that are close to each other and contain continuous visual information such as brightness and hue. For example, if a neural network sees a flower in one pixel, it recognizes it in the pixels standing next to it. The structure of this neural network is divided into two types of layers — convolutional and pooling (Fig.1). In convolution, the neural network removes unnecessary information and leaves only the necessary information that will help in recognizing the image. A convolution can be created for any feature, which is selected by the network itself. After that comes the pooling layer, which helps to reduce the size of feature maps, which reduces the load on the computer for further computations [2].

At the hardware level, the camera operation starts with a lens that collects light from the observed scene and focuses it on the image sensor. The sensor, usually of CMOS or CCD type, converts the light into an electrical signal. The quality of the lens and sensor directly affect the clarity, sharpness, and viewing angle of the image. The main computational load falls on the processor, which is the heart of the camera. The processor typically consists of a CPU, which performs the main computational tasks and controls peripherals, a GPU, optimized for parallel processing, which accelerates the execution of AI algorithms, and an NPU/AI Accelerator, a specialized processor that accelerates neural networks. Also, the hardware includes RAM (RAM) for temporary data storage, flash memory for storing firmware and AI models, an SD card slot for local video storage, a network interface (Ethernet, Wi-Fi) for data transfer and remote control, I/O interfaces for connecting external devices (microphone, speaker), and a power supply (Fig. 2). In addition, cameras can include infrared illumination for operation in the dark. At the software level, the camera operation is provided by the firmware, which is the camera operating system and a set of programs that control the operation of all hardware components. It includes device drivers, network protocols, and APIs for developers [3].

Figure 2. IP surveillance camera device

A key feature of AI cameras is local data processing, which reduces network and server load and enables faster system operation. Intelligent algorithms that use neural networks for complex analysis and recognition tasks, as well as automation that provides self-detection and response to events, play an important role. Adaptability, which provides the ability to retrain and customize algorithms for specific tasks, makes the system more flexible and efficient.

Object detection is a fundamental process in AI video surveillance systems that allows the camera to understand what exactly is happening in the frame, rather than just recording the video stream. This process is the first step in more complex tasks such as recognizing, tracking and analyzing object behavior. The object detection method involves determining the presence and location of objects of a particular class. The object detection process begins with data collection and preparation. This step includes data partitioning, where images are manually or semi-automatically marked up with selected and signed objects, and image preprocessing, which includes image resizing, pixel value normalization, and data augmentation (Fig. 3). Augmentation, i.e., artificially increasing the amount of data by applying various transformations, increases the diversity and quality of the training sample [4].

Figure 3. Object detection process

The next step is the selection of the neural network architecture. Converged neural networks are the basis of most modern detection algorithms. The architecture selection is followed by the stage of training the model with a teacher, when the model adjusts its parameters to minimize the error between the predicted and actual positions of objects in the images. In the detection phase, after training, the model analyzes the input image and returns the coordinates of the bounding boxes and the classes of detected objects. The application of the object detection method finds a wide range of applications in video surveillance systems with AI, ranging from security and safety, where it is used for intrusion detection and access control, to monitoring production processes in industry, traffic analysis in transportation, counting visitors in retail and monitoring public space in the smart city concept [5].

The application of artificial intelligence in video surveillance systems opens up new opportunities to improve the efficiency and security of various areas. Intelligent video cameras equipped with powerful image processing algorithms and neural networks allow automating the process of video stream analysis, reducing the time spent on searching and analyzing information. This leads to a significant increase in the accuracy of event detection, reduction in the number of false alarms and optimization of resources required for storage and processing of video data. Object detection, facial recognition, behavioral analysis — all of these capabilities increase the efficiency of security systems, streamline processes and improve service delivery. However, despite the impressive progress in AI for video surveillance, there are challenges. The cost of such systems, as well as the need for skilled personnel to set up and operate them, are still important factors. Issues of privacy and ethical use of facial recognition technologies also require careful consideration and the development of appropriate regulations.

Further development should focus on reducing the cost, improving the reliability and robustness of algorithms, and developing ethically acceptable and safe methods of applying AI in surveillance systems. Research and development in video processing algorithms, as well as computational platforms that enable complex real-time computations to be performed on camera hardware, remain critical to further progress in this area. The integration of video surveillance systems with other systems, such as access control systems or public address systems, also represents a promising direction for the development and efficiency of intelligent video surveillance systems. It is these steps that will expand the applications of this technology and make it more accessible and effective for a wide range of applications.