Seeing doesn’t mean understanding – security must also be "smart"!

Just ten years ago, “security” was more or less a guard with a dog at most factories and warehouses. As machines gradually replace manpower, a large number of cameras, network video recorders (NVR), infrared detectors and other pieces of equipment have been incorporated in both public and private security systems. Nevertheless, machines are not the be-all and end-all of security. The flaws of so-called “smart’ security systems have been extensively portrayed in movies, and their vulnerabilities have been exposed in real life by enterprising hackers who disarm them with a pair of pliers or a few pieces of chewing gum. Since we now live in a society that wants every single device to be smart, security needs to learn from its past mistakes and institute its own smart upgrades.

Pain points of traditional security

“Overwhelmed” by numbers

In public spaces in many cities, dense masses of people and large fluctuations in population are major issues for monitoring systems, which struggle to extend full coverage and accurate positioning. In addition, the occurrence of large conferences, events and holidays cause local populations to temporarily swell, posing further challenges to security monitoring.

Passive "defense"
Traditional monitoring systems typically address problems after they have occurred, and are simply unable to muster an offense or prevent an attack. In most cases, security personnel are only able to review video stored in a recorder after an incident has occurred, which means devoting many person-hours to analyzing it second by second, frame by frame. Think of the ease of getting instant answers via search engines, then imagine what it would be like look up a word or line of text in a book without an index or table of content. That’s an apt analogy for the tedious and totally inefficient process of reviewing security footage in search of irregularities that could shed light on a situation after the event.

Swallows up resources
Large amounts of video data and low utilization are also pertinent reasons for the urgent need to “smartify” traditional security. The global total of data generated by video surveillance was about 18.1 PB (1 Petabyte=1024 Terabytes) in 2020 alone. This massive amount of data accounts for 83.1% of the IoT data over the same period. In practical applications, however, various labor and technical constraints keep the utilization and efficiency of data very low. Setting up a basic video surveillance network and infrastructure is not difficult these days. The difficulty lies in using data to provide both rapid and accurate support for security businesses.

Isolation of information
Data islands are unavoidable issues in the overall building of smart cities. They also constitute a problem that needs to be addressed in the smart upgrading of security. In traditional security systems, gaps exist among jurisdictions, platforms and systems that are difficult to bridge, be it hardware matching or platform architecture. In scale networking and smart transformation, the difficulty and cost of implementation have become problematic issues.

Limitations inherent in traditional security technologies are becoming increasingly obvious, and various industries have varying requirements for security technology application scenarios. Benefiting from new technologies such as AI and IoT, the security industry has also welcomed the opportunity to join hands with AI as well as machine vision (MV) technologies. Innovations such as video recognition and facial recognition have found applications in security scenarios. Especially after the pandemic, a new rush of demand for facial recognition has emerged in places like parks and office buildings. At the same time, the pandemic has accelerated the pace and the focus of the development of smart security. AI-supported security has become a dedicated direction for many technical fields.

Sensor upgrades

People usually compare camera lenses with human eyesight. This "expectation" is still there even today, and still provides a reference-point for the smart upgrading of cameras. Eyes, first and foremost, must be able to see and see clearly. Not surprisingly, we have witnessed continuous upgrades in video resolution, from 720P to 1080P, 2K, and 4K. However, high resolution is not the only mandatory requirement for "seeing clearly". Once environmental conditions fail to meet requirements, these numbers also become empty standard parameters. For the camera lenses to “see clearly”, the contact image sensors (CIS) of the camera must have a complete series of intelligent perception functions.

In security applications, requirements for image clarity and scene coverage continue to increase. In order to provide clear details and color fidelity image information during daytime hours with adequate lighting, complex light environments – such as morning, dusk, and night, will carry more stringent requirements for performance of night vision in the CIS. Therefore, very-low-light imaging, product performance, color expression, and near-infrared (near-IR) imaging performance of the image sensor have also become basic technical conditions for the implementation of smart security.

Machine vision

A major reason for low efficiency in traditional security systems is that the recorder stores images taken by the camera in an almost "mindless" way. Although resolution is constantly improving and storage space is expanding, the information obtained is little more than "high-quality", seldom-used data. To improve efficiency, the system needs a layer of intelligent analysis. Assisted by the computing power of the analysis system, large amounts of data can be analyzed, filtered and processed in order to efficiently extract the information required.

MV technology applications have introduced new changes to the security industry. The main goal of MV is to give the computer the ability to recognize 3D environment information from two-dimensional images, and then process such geometric information as shape, position, posture and movement of objects in that environment. Advances in target recognition, target tracking, stereo technology and multiple dome camera tracking linkage technology have also occurred in the same industry. Analytic functions provided by MV now allow for the possibility of security systems to "understand". For example, it can be combined with image processing technology in the design of a real-time monitoring system that can monitor and record while adding detection of image signal changes and automatic recording functions through MV technology. This allows the system to simultaneously identify scenes and issue alarms and thus realize a fundamental improvement – from passive to active.

Real-time processing of data

In the process of AI data analysis, real-time and near-real-time data processing capabilities are indispensable. This often requires the system to start AI reasoning and recognition on smart edge platforms. Currently, these analytic and reasoning processes are mainly carried out on the image/vision processor and image signal processor chips. The "on-site processing" of data requires the various modules to operate in coordination. First and foremost, valid data needs to pass through the image signal processor (ISP) to obtain clear image data before using the neural processing unit (NPU) to perform real-time calculations. At this point in time, the amount of computing power on the end side determines how many functions it can accomplish. For example, when the end side computing power reaches 1.5T, it can meet the needs for running 3 to 5 algorithms simultaneously and perform functions such as facial detection, recognition, and tracking. The System on a Chip (SoC) requires strong integration capabilities in terms of performance, including modules such as ISP, NPU, and signal codecs, to provide sufficient AI computing power on the user end.

Currently, most edge computing on the market is oriented toward 4-16-channel signal analysis processing, such as in vehicle-road collaboration and gas stations, or supporting 200-channel small-scale data centers, as seen in oil production plants, transformer substations and other scenarios. In such scenarios, real-time data processing greatly bolsters data privacy protection for the user, while also assisting on the cost- and energy-saving aspects.

High computing C/P value

The whole point of customization is improved efficiency. This truth applies to all industries. In AI computing for smart security, dedicated AI vision chips are simply more capable than general-purpose CPU/GPUs at achieving better computing performance, efficiency and cost-effectiveness.
Take domestic AI signal chips, for example. In a data center using 100 AI servers to process 25,000 signal channels for analysis, if an AI-dedicated and more cost-effective data processor AI chip is used, it can reach an actual computing power that is more than four times superior in performance. In other words, for the same application with the same performance, the number of 100 AI servers once needed is reduced to only 25. This translates into cost savings of more than 70% for the data center. This financial fringe benefit ensures that smart security has great commercial potential.

Demand for "AI+Security" in B-end corporations and C-end personal security markets has gradually expanded over recent years. And we have data to prove it. In 2020, the scale of the AI+security software and hardware market in China was 45.3 billion RMB, and it is expected to reach 54.2 billion RMB in 2021 – a year-on-year increase of 19.5%. In this new race, players are not only pitting their “security guard” skills against each other, but also racing to be the first to achieve cross-category, cross-industry integration of multiple technologies.

Perhaps the smart security that will outsmart all others will be the one that is most “comprehensive”.