“In recent years, the size of the edge computing market has continued to grow, and it has been rapidly applied in security, smart city, autonomous driving and other fields. The processing, analysis and storage capabilities of the edge have become the key to testing AI companies.
In recent years, the size of the edge computing market has continued to grow, and it has been rapidly applied in security, smart city, autonomous driving and other fields. The processing, analysis and storage capabilities of the edge have become the key to testing AI companies.
Zhao Hanwei believes that the Internet of Things and the cloud have been developed for many years, and the completion of platform construction and connection has been very high, but the smart city has not really arrived because the basic perception has not been realized. Chip computing power and edge computing determine the perception ability. Only by intelligent analysis at the front end and cooperation with the back end can the urban IoT data be fully utilized to create a perception city. Touchview has been focusing on the use of artificial intelligence technology to perform computing at the edge to realize the field of edge intelligent perception. Over the past ten years, it has accumulated a large number of practical cases in various sub-fields. Zhao Hanwei combined Touchview’s application in subway scenes to interpret the edge Problems and unique solutions encountered by the computing side in practical scenarios.
The following is an excerpt of the content shared by Zhao Hanwei.
Touch View Infinite and Edge Computing
The principle of edge computing is similar to that of an octopus. An octopus has 40% of its neurons in the brain and 60% of its neurons in its tentacles. The eight tentacles can sense contact information and even do simple thinking. The tentacles do part of the brain’s work, reducing the workload of the brain’s processing. Edge computing is distributed computing. Part of the calculation is done at the front end close to the sensor, and the calculation results are directly fed back to the front end and the brain.
Most of the architectures are now combined with the cloud and the edge, so the edge must have fast processing capabilities. For example, the sound sensor turns the camera to the angle of the sound source after hearing the sound, which is actually a quick response after processing at the edge. If the sound command is sent to the cloud in the center, the cloud calculates and then sends the command to the front end. At this time The sound landscape may have changed, so fast response at the front end is an advantage of edge computing. In addition, many scenarios currently cannot guarantee the real-time connectivity and real-time speed of the network. Although the most typical autonomous driving has a cloud-based brain, it is difficult for the car to fully guarantee the signal while traveling, so there are corresponding edge computing devices in the car to process it. Perceive the problems encountered by the front end.
Cloud AI and embedded AI are not opposing technologies, and the combination of the two can meet most needs. In the cloud, there are abundant computing resources, fast acceleration, and can support very complex models and algorithms. At the same time, a real-time networking environment must be guaranteed; on the edge, they are basically embedded, with limited computing power. And compression, applicable to a variety of scenarios. The cloud and the edge work together. Touchview has ten years of experience in edge computing. First of all, in the accumulation of perception algorithms and intelligent recognition algorithms, the algorithms not only include video structuring, sound processing, various sensor processing, but also model compression and model optimization.
For example, let everyone understand the algorithm compression and optimization. An AI state-owned enterprise that cooperated with us recently can compress the algorithm to 50M at most, and the efficiency of recognition will decrease if we continue to compress it. The algorithm we execute at the edge can be compressed to hundreds of K. to 1M, and the recognition efficiency does not decrease under high compression. This is a manifestation of our long-term accumulation of edge algorithms. In addition to understanding the characteristics of specific chip application scenarios, we also have the ability to fully integrate the algorithm and chip with the underlying optimization to fully tap the computing power of the chip. For example, the Intel Movidius2450 chip has 200G computing power. By injecting assembly language at the bottom of the chip algorithm, we change its data preparation, calculation, etc., in the actual measurement, it can be comparable to the 1T computing power chip on the market. The technical content of this capability is actually very high. high.
We have injected edge computing technology and accumulated experience into the company’s five series of products. The edge computing unit is actually an edge computing device. First, it can be connected to IoT devices, and secondly, it can process the connected IoT device signals at the edge, and then convert them into perceptible things. It can also do some For front-end control, these data are transmitted to the cloud through a secure encryption method, which is the function of the edge computing unit. In this epidemic, we quickly launched a series of instantaneous body temperature screening, which uses infrared imaging and visible light imaging to perform superimposed operations to achieve rapid body temperature screening. For the most advanced products, we use a 640*480 infrared sensor chip, which can be passed in 1 minute. The number of temperature measurement can reach 780 people, and the performance is very high in the entire temperature measurement market.
The Dunwu series is mainly for the stock market. For example, high-definition cameras do not have intelligent functions. After adding this box, ordinary cameras have face recognition, human body, and vehicle capture functions. Horned Hummingbird is a development kit based on artificial intelligence that we cooperate with Intel, mainly artificial intelligence research and development companies and education and training institutions, to give students artificial intelligence training courses, plug in the computer through USB, and 20 lines of code to get face recognition, this It can detect 20 kinds of physical models, can do rapid artificial intelligence development, of course, can also develop models, replace the models inside and add new algorithms.
We work with more than 2000 companies and educational institutions. Snapshot series cooperates with camera manufacturers to provide smart cameras to the market, adding an edge computing component to the camera, with algorithm models and chips to turn the camera into a smart camera. One of the main scenarios is as an Electronic police, which can detect at intersections 14-18 kinds of illegal behaviors, including behavior analysis, vehicle detection, personnel detection, etc.
Perceiving the understanding of the city
The Internet of Things and the cloud have been developed for many years, and the construction and connection of various platforms have been almost completed. Why is the smart city still not coming, because the basic perception has not been realized. The development of perception technology is actually related to the Internet of Things technology, edge perception technology, and cloud computing technology. The Internet of Things and cloud computing appeared more than ten years ago, but perception has not been promoted rapidly, mainly because the chip computing power is not enough, dragging wisdom. The hind legs of the city. Smart city construction is inseparable from intelligent perception, including water, energy, transportation, green space, finance, etc., are inseparable from sensing technology. With sensors, intelligent identification is required, and intelligent control is required after identification. Therefore, in the entire system , perception is a very important part.
In recent years, with the increase of chip computing power, edge computing has become a very important topic. In 2018, MIT Technology Review listed sensory cities among the top ten breakthrough technologies in the world. In the past two years, urban digitization has become our national strategy. To digitize a city in the digital economy requires sensors to upload various data in the 3D model of the city to form a true digital city. Many data are difficult to understand, so it needs to be on the edge of the chip. Calculation, the realization of what is transmitted to the central end, is already something that can be read.
Project case sharing
The case shared today is the subway face entry. In the past, there were basically two access schemes for urban subways, swiping cards and QR codes. Cards are easy to be lost, forgotten, damaged, and troublesome to return and exchange cards; QR codes provide convenience for everyone, but during peak times, the network is congested, and the QR codes cannot be opened, causing congestion and greatly affecting travel efficiency. This is a problem with the current subway traffic plan. Many people discuss the plan of swiping your face to enter the subway. In fact, there are a few issues that need to be considered.
Can the speed of travel be guaranteed? Will it be misidentified? swiped the wrong debit account? Will the number of users affect the processing power? Is the capital investment large?
The traffic speed should be guaranteed to be around 200 milliseconds. At present, it is temporarily impossible to realize the “automatic opening of the gate from the snapshot to the person”. Because at present it is by installing a small screen reader on the gate. Basically, it is a semi-coordinated type, and passengers have to pause a little to increase the speed of traffic. Misrecognition is the category of face technology, and the current technology is difficult to achieve 100%, which is difficult. A recognition rate of 99% is not bad. The standard of the Ministry of Public Security is above 95%, but 95% is unbearable for the subway. The daily flow of people on the subway in a medium-sized city is about 3 million. How many people have a 95% misrecognition rate of 3 million people? Therefore, some auxiliary means should be used to make up for the inherent shortcomings of the face coefficient. For the problem of large number of people, consider using the method of pre-screening and grading lists, which is called shrinking in the subway industry.
For example, 10 million people in the whole city have registered their faces, but the actual number of people entering the station is not even 10,000, and the total number may be 200,000 a day. At this time, people entering the subway station need to be screened. In a city like Beijing, the number of face database registrations will reach tens of millions. If you search for a photo of a person in the face database of tens of millions, you may not be able to find it in at least 10 minutes, and the subway company is likely to be reluctant to invest. Therefore, we need to solve technical problems through practical means.
This plan has four goals. One is to build a subway face-swiping payment system to improve management and control capabilities; on this basis, to improve the travel experience of passengers, such as directly scanning the face to enter the station to solve the problem of inconvenient to take a mobile phone; in addition, it also improves the public security capabilities. Then precipitate big data, conduct data management and customer behavior analysis on desensitized data, and obtain more value from the data. Of course this information is confidential.
This project has three technical key points.
First, face selection technology. The face captured by the lens is tracked, the camera fills the dark face area with light, and the side face with poor contrast effect is filtered out, and the best shot is selected in the whole process.
Second, face pre-screening. During the peak traffic hours in the subway, there may be hundreds of people in one shot. If they go to the gate and then search the face database of tens of millions of people, they still need to wait. There is generally at least a few tens of seconds between the subway passage and the turnstile. In fact, the cloud face can be retrieved in advance, and the retrieval results can be placed on the turnstile. When the passenger walks to the turnstile, they can quickly identify and pass.
The third is the local library intelligent screening. Build a local frequent visitor database based on big data analysis technology. For example, if someone enters and exits station B from station A every day, they can cache this information at the corresponding site; you can also build a VIP library.
How is the whole program realized? The passenger terminal downloads the mobile APP, and after registration, the cloud system collects information such as faces; then binds the payment information, connects to the billing system, and the cloud synchronizes the data to the subway face database. Select the best face in the face database for two-stage pre-screening, and then quickly download the screening information to the gate terminal. This process lasts up to 2 minutes.
The gate will open the door after a quick comparison. All the information of people entering the station will be put into the face database of the station. When leaving the station, the same will be pre-screened in the station, and the comparison will be done in the warehouse of all the people entering the station. After the comparison, it will start at the gate. At this time, the passenger entry and exit records have been sent to the billing system, and the connected electronic wallet will automatically deduct the fee. This is the overall program composition. This plan has brought advanced AI face recognition technology to the subway. From the perspective of the Sino-US trade war, the country has adopted technologies such as artificial intelligence and big data as national strategic technologies, so it is not a question of whether or not to use these technologies, but we need to Through the upgrading of these technologies, the development of the entire industry and industrial chain is driven.
The real-name system for passengers has been realized by brushing their faces into the station. An important role here is the bank. After the bank obtains the face information, the real-name comparison is performed in its face database to provide big-data personnel portraits. After capturing a large number of faces, based on the images of the faces, the attributes such as gender and age group can be roughly analyzed. With this information, for banks and subway operating companies, statistical analysis and decision-making analysis are based on these desensitized data; In addition, the overall security capability has also been improved. In some critical situations, such as catching criminals, you can go to the information database to adjust the corresponding data, and see the person’s movement trajectory, entry and exit sites and time. It can be said that we have basically achieved the goals of the program.
For example, for a medium-sized city with 128 subway stations, the average daily flow of people in the subway is about tens of thousands, and about 3 million people during peak hours. Our solution, each station has 4 station entrances, and each station entrance has two capture cameras. , 8 video-structured terminals, and a video recognition server, which can be recognized directly at the station. We finally tested the speed of 53 people per minute, which is much faster than swiping cards and QR codes. This is the technical ability of Touch View Infinite accumulated in the past 10 years. The close combination of algorithms and chips has deeply tapped the computing potential of the chip. The performance of the same chip and the same model far exceeds the industry standard level, and the products are more competitive.