AI adds another lock to security

Introduction

In urban public places, dense crowds and large population flows are a major challenge for monitoring systems. It is difficult to achieve full coverage and accurate positioning in these areas. In addition, during some large-scale conferences, events or holidays, the population flow in the area will increase instantly, which will also increase the difficulty of implementing security monitoring.

Traditional monitoring systems tend to respond after the fact, lacking the ability to proactively attack and prevent beforehand. Staff can usually only retrieve the video stored in the surveillance recorder after the fact, and use a lot of manpower to analyze it second by second and frame by frame, which is like looking for a few lines of text in a paper book without a table of contents. For those of us who are used to using search engines, the efficiency and difficulty of such work can be imagined.

The large amount of video resource data and low utilization rate are also the main reasons why traditional security urgently needs to be intelligent. In 2020, the data generated by global video surveillance is about 18.1PB (1PB=1024TB). Such a huge amount of data accounts for 83.1% of the IoT data in the same period. However, in actual applications, due to various human and technical conditions, the utilization efficiency of these data is still very low. Nowadays, it is not difficult to build a basic video surveillance network and infrastructure. The difficulty lies in how to use these data to provide fast and accurate support for security business.

Data silos are an unavoidable issue in the construction of smart cities, and are also a problem that needs to be overcome in the intelligent upgrade of security. In the traditional security system, there are difficult-to-fill gaps between various jurisdictions and platform systems, whether in hardware matching or platform architecture. In large-scale networking and intelligent transformation, the difficulty and cost of project implementation have become thorny issues.

People usually use eyes to compare with cameras. Although this "requirement" still seems very high now, it also provides a reference direction for the intelligent upgrade of cameras. The function of the eyes is first to be able to see and see clearly. We have also witnessed the continuous upgrade of video resolution from 720P to 1080P, 2K, and 4K. However, high resolution is not a necessary and sufficient condition for "seeing clearly". Once the environmental conditions cannot meet the requirements, these numbers become false parameters. In order for the camera to "see clearly" more actively, the camera's image sensor (CIS) needs to complete a series of intelligent perception functions.

In security applications, the requirements for imaging clarity and scene coverage will continue to increase. In addition to providing detailed and realistic image information during the day with good lighting, the requirements for CIS night vision performance in complex lighting environments such as dawn and dusk and at night are also more stringent. Therefore, the low-light imaging, product performance, color expression and near-infrared imaging performance of image sensors have also become the basic technical conditions for the implementation of intelligent security.

One of the main reasons for the low efficiency of traditional security systems is that the recorders store the images taken by the cameras in an almost "brainless" way. Although the resolution of the images is constantly improving and the storage space is constantly expanding, what is obtained is more "high-quality" invalid data. To improve efficiency, it is necessary to add a layer of intelligent analysis to the system, using the computing power of the analysis system to analyze, filter and process the massive data to extract the information people need.

The application of machine vision technology has brought new changes to the security industry. The main goal of machine vision is to enable computers to recognize three-dimensional environmental information through two-dimensional images, so as to process geometric information such as the shape, position, posture, and movement of objects in a three-dimensional environment. Under its category, target recognition, target tracking, binocular technology, and multi-ball tracking linkage technology have also been derived. The analysis function provided by machine vision makes it possible for security systems to "understand". For example, it can be combined with image processing technology to design a real-time monitoring system. While monitoring and recording, the system can use machine vision technology to add video change detection and automatic recording functions, and can simultaneously perform functions such as scene recognition and alarm, achieving a fundamental improvement from passive to active.

In the process of AI analysis of data, real-time and near-real-time data processing capabilities are essential, which often requires the system to start AI reasoning and recognition on the intelligent edge platform. At present, these analysis and reasoning processes are mainly carried out in image/visual processors and video processor chips. To achieve "on-site processing" of data, it is necessary for various modules to be able to operate in coordination. First, valid data must pass through the ISP to obtain clear image data, and then the NPU (neural network processor) will calculate the data in real time. At this time, the size of the computing power on the end side determines how many functions it can achieve. For example, when the computing power on the end side reaches 1.5T, it can meet the needs of running 3 to 5 algorithms at the same time, and perform functions such as face detection, recognition, and tracking. In order to enable the end side to have sufficient AI computing power, the SoC needs to have strong integration capabilities in terms of performance, including ISP, NPU, video codec and other modules.

At present, edge computing in the market is mostly aimed at 4-16-channel video analysis and processing, such as vehicle-road collaboration, gas stations, etc., or supports small data centers with about 200 channels, such as oil production plants, substations, etc. In these scenarios, the real-time data processing capability is of great help in protecting user data privacy and saving costs and energy.

Professionals do professional work, which is a truth applicable to all industries. In the AI ​​computing of smart security, the use of specialized AI vision chips can achieve better computing performance, high efficiency, low cost and many other benefits compared to general CPU/GPU.

Taking a domestic AI vision chip as an example, in a data center scenario where 100 AI servers are used to process 25,000 video analyses, if a data stream AI chip dedicated to AI and with a higher cost-effectiveness in computing power is used, the actual computing power can be more than 4 times higher. This means that for the same application to achieve the same performance, the number of AI servers required has been reduced from 100 to only 25, which has brought more than 70% cost reduction to the data center. This cost-effective "economic account" also ensures that smart security is not only an upgrade in demand, but also has great potential in business.

End of article interaction