Three major development trends of AI in the security industry

According to market estimates, by 2026, global AI hardware, software and services spending will exceed $300 billion, with a compound annual growth rate (CAGR) of 26.5% from 2022 to 2026; The market size of security applications combined with AI technology will also increase by more than 10 billion US dollars, with a CAGR of 18.64%, and the growth rate is increasing.
Therefore, security manufacturers are scrambling to enter the embrace of AI, especially image-based monitoring manufacturers, AI technology has long been applied to the detection, search, tracking of monitoring images for specific objects, and can automatically trigger, linkage alarm and other related system equipment, greatly reducing human and material resources, and improve the efficiency of security management.
Although AI is good to use, but the hardware equipment and energy-consuming load that support its calculation are quite high, so in the pursuit of accuracy, how to reduce the load, make AI more “light”, faster, and solve the problem of the general lack of AI professionals in the enterprise, reduce the cost required has become the direction of the industry’s joint efforts. At this year’s Secutech exhibition, a&s observed three major trends in the development of AI in the security industry:

Number one: Cloud to edge
In recent years, it has become more and more obvious that AI computing capabilities are gradually developing from back-end/cloud to Edge (Edge), the main benefits are reduced time, reduced risk and energy consumption, and of course, the most important – cost savings. For example, the Edge AI solution proposed by IT giant Intel emphasizes that only a general PC, Notebook or x86 industrial computer is needed, and through the OpenVINO open platform, more than 200 algorithms on the platform can be used (including TensorFlow, CaffeZ, etc.). Develop or use AI models that have been pre-trained by third-party partners to easily transform terminal devices into Video AI boxes, NVR with AI functions or Edge side AI servers, thereby solving problems faced in various fields and making AI systems no longer costly.

Edge AI solutions launched by security manufacturers are also quite diverse, especially surveillance camera manufacturers have put AI video identification, analysis, search, tracking and other functions in the front end. For example, Vivotek uses Edge-centric object extraction on the front end to perform analytical operations without installing high-performance graphics cards on the back end, effectively reducing server computing resources and construction costs. LILIN has long seen that Edge AI will become the mainstream, and is committed to the application of edge computing AI cameras combined with 5G and cloud. And launched the latest AI waste Detection (Trash Detection), Ball Tracking (Ball Tracking), Gender detection (Gender Detection) and other functions, its video management software (VMS) is designed for edge computing AI camera integration. The AI video identification edge operation developed by AiUnion, a software manufacturer, integrates Image classification, object detection and Image Segmentation technologies of AI deep learning to establish a general model such as smart industry, science and technology law enforcement, security monitoring, etc., which is easy for users to directly import and use. If there is a specific application and then tailored to its needs.

Number two: Go from heavy to light
Generally speaking, AI will re-identify similar images (such as background) as independent images when identifying videos, so the calculation is large; When there is a large field to do AI video identification, the number and cost of hardware equipment to support its calculation will be expensive, and the relative energy consumption is high, which does not meet the principle of enterprise sustainable development (ESG). Therefore, many manufacturers have tried to find ways to “do subtraction” for the current AI operation mode, so that they can wave away the heavy burden, show a light body and take a light step.

for example, electronics giant DELTA has introduced DIVA (DNN Inference OS for Video Analysis), an intelligent accelerator that uses the properties of a series of similar images to speed up AI video analysis, which can be applied to static or dynamic cameras. As long as a layer of DIVA SDK software is added to the AI model of any imaging application, it can accelerate 2 to 6 times (different field of view hardware, video resolution, AI model) without losing any accuracy, thereby reducing the number of devices and energy consumption to achieve cost savings.

Nexuni, a startup, also proposes the use of low-power, miniaturized, highly customized Embedded systems (such as Embedded Linux, Nvidia Jetson Platform, etc.). In combination with the machine learning model TinyML, which can be downgraded to the KB storage level, it becomes the Edge AI system architecture – it can perform AI operations on video (such as faces) and voice on the hardware specifications of the endpoint with limited resources, and maintain the same accuracy as Server level operations. According to the actual measurement, TinyML can reduce the 16GB required for Server level to 320KB for edge embedded systems, which is equivalent to reducing the traditional machine learning model by 50,000 times, making it easier to develop automation solutions that can be applied to various specifications systems and easy to popularize, and improve management efficiency.

Number Three: from professional to universal
ChatGPT has brought a lot of attention to AI around the world lately, mainly because it makes it easy for people to use AI. Similarly, in order to accelerate the adoption of AI in the market, the industry is also trying to simplify the development technology and design process, hoping that even people without AI expertise can easily create effective AI training models that meet their own business needs. Intel’s Geti computer vision platform, for example, is touted as enabling anyone to easily deploy high-quality computer vision AI through a simple data upload, tagging, model training, and retraining interface optimized with the OpenVINO toolset to drive more application innovation and improve overall enterprise performance.

Considering that the general AI project import process: data preprocessing → selected algorithm → program development AI model → model validation, not only requires AI or IT professionals, but also requires repeated operations, which can take months; AutoML Platfrom (No-Code AI platform), developed by Profet AI, uses machine learning technology to automatically and quickly build AI models in two steps: “Select algorithms → program development AI models” and optimize model verification. In this way, various departments of the enterprise (such as human resources, research and development, production, IT… People can easily apply AI to solve various problems in their business fields. For example, AutoML Platfrom can be applied to industrial production, which can optimize the manufacturing process and improve the overall yield and efficiency. Applied to network security, it can also help predict the degree of security and prevent the leakage of risky emails.