Urban hot security market development hotspot in 2018 and development prospects in 2019
Jan 28, 2019

The development hotspot of the video surveillance market in 2018 is that AI gives "noisy". Throughout 2019, data will lead the development of the video surveillance market.   First, AI empowerment   Alibaba Group Vice President Zeng Ming pointed out in the preface to the book "Redefining the company: how Google operates", "Although the future organization will evolve into what it is, it is still difficult to see, but the most important function of the organization in the future has been It’s becoming clearer that it’s empowerment, not management or motivation.” Others say it’s one of Google’s founders, Larry Page.   "Energy" is to give the object some power and energy. Generally speaking, you can't, but I can make you. It was originally a vocabulary in psychology, designed to give positive energy to others through changes in words and deeds, attitudes, and the environment. The term is now widely used in various industries.   AI empowerment is to assign AI's ability to third parties. We can say that AI can empower cities, AI enable security, AI enable video surveillance. Today we explore a small concept AI how to enable video surveillance?   The function of the traditional security system now seems a bit too simple, and we don't even know we own a gold mine until AI is not mass-marketed. Taking video surveillance systems as an example, traditional video surveillance is real-time monitoring, recording, and playback of video. Over the set time and then overwriting the old video with new video, this phenomenon lasted for about 60 years until 2016, AI Fu When you enable video surveillance, computer gradually began to be able to read one frame of the picture, there is three big AI technology is the core of human identification, vehicle identification and Reid .   ◦ Vehicle identification : Vehicle identification includes two major technologies: license plate recognition and vehicle feature recognition. The license plate recognition technology was first assigned to the video surveillance system, and the multi-application was applied to the card-free entrance and exit management system of the bayonet, electronic police and parking lot. License plate recognition belongs to the category of OCR text recognition in a certain sense. The only difference is dynamic license plate recognition. At present, the vehicle feature recognition can be more than 20 kinds. It can be said that the potential of video and image is greatly explored, and the logo, color and marker are relatively classified and easy to implement, so the market has appeared. Many vehicle big data platforms have also appeared in various vehicle technical methods and applications, which are the result of AI empowerment.   ◦ Human body recognition : Human body recognition includes two technologies : face recognition and body feature recognition. Face recognition is much more complicated than license plate recognition, and is further subdivided into a combination of (such as access control) and non-fitted modes (such as open environment acquisition), especially non-cooperating dynamic face recognition technology in 2017. Only a significant increase in the recognition rate of more than 70% into commercial use, and it is precisely face recognition technology that empowers the entire security industry, after all, the core of security system management is people (the other core is the car), once the identity of the person is identified The rest of the out will be handled more. Human body feature recognition is an accessory of face recognition. The face can be used to judge gender, age, skin color, whether to wear the eye, and the whole range can be recognized by enlarging the recognition range. The identification includes the color of the shirt, the color of the shirt, whether to play the umbrella, whether or not For example, the body recognition technology has more power for video surveillance systems than vehicle identification.   ◦ ReID : There is a situation where the monitoring system can't see the face or can't see the face. This depends on the pedestrian re-identification technology (ReID). The author can assert that ReID must be the development of future video surveillance. I also believe that at this point. ReID technology does not require a special camera, and the environment requirements are not so high. As long as certain pedestrian characteristics are identified, pedestrian trajectory analysis can be realized, and cross-mirror tracking can be further realized. Once a face appears on the trajectory Then, the identity of the entire trajectory can be clarified. This is the best technology for law and order. After all, we have installed so many public security cameras.   Second, AI in 2018   After discussing AI empowerment, let us explore AI. After all, AI is the core hotspot of 2018. In my understanding of AI, I intend to use some of the information on the network to summarize, by the way, what will happen in 2019.   2.1 Natural Language Processing (NLP)   The watershed in the history of NLP in 2018. In 2018, breakthroughs in the NLP field continued: ULMFiT, ELMo and the recent hot BERT, which of course is not entirely related to video surveillance. Migration learning has become an important driving force for NLP progress, starting with a pre-training model and constantly adapting to new data, bringing endless potential.   According to the 2019 authoritative outlook of ULMFiT author Sebastian Ruder, it is expected that “pre-training language model embedding will be ubiquitous, without pre-training models, training from the beginning to the top level of the model will be very rare. Pre-training representations that can encode professional information will It will appear, this is a complement to the language model embedding. At that time, we can combine different types of pre-training representations according to the needs of the task. There will be more research on multi-language applications and cross-language models. On the basis of cross-language word embedding, deep pre-training cross-language representation will appear."   2.2 Computer Vision (CV)   In 2018, a large number of new researches have appeared in both image and video directions. Three major studies have set off a collective wave in the CV circle. It has also been said that the biggest progress in 2018 is that there is no progress.   ◦ BigGAN: In September 2018, when the ICLR 2019 paper in the double-blind review of BigGAN appeared, the experts were boiling: it was hard to see that this was generated by GAN himself. In the history of computer image research, the effect of BigGAN has improved a lot. For example, after performing 128×128 resolution training on ImageNet, its Inception Score (IS) score is 166.3, which is the best score of 52.52 points and 3 times. In addition to the 128×128 thumbnails, BigGAN can also train directly on 256×256, 512×512 ImageNet data to generate more convincing samples.   ◦ : Train the entire ImageNet in 18 minutes. In August 2018, Jeremy Howard, the founder of the online in- depth learning course, and his students used the Amazon AWS cloud computing resources rented to train the image classification model on ImageNet to an accuracy of 93% in 18 minutes.   ◦vid2vid technology : In August 2018, NVIDIA and MIT's research team produced an ultra-realistic HD video generation AI. As long as you have a dynamic semantic map, you get a video that is almost identical to the real world. In other words, as long as you sketch out the scenes in your heart, you don't need real shots, and movie-level videos can be automatically popped out. In 2019, in the field of computer vision, the improvement and enhancement of existing methods may be more than the creation of new methods. The self-supervised learning of this year's fire may be applied to more research next year.   2.3 Tools and frameworks   AI empowerment is inseparable from tools and frameworks, and tools and frameworks in the machine learning arena are still evolving rapidly:   ◦PyTorch 1.0: According to the 2018 annual report released by GitHub in October, PyTorch ranked second in the fastest growing open source project. It is also the only deep learning framework for finalists. As the benchmark of TensorFlow, PyTorch is actually a recruit, officially released on January 19, 2017. In May 2018, PyTorch and Caffe2 were integrated to become the next-generation PyTorch 1.0, and the competitiveness was further strengthened.   ◦AutoML: AutoML is a new way of deep learning that completely changes the entire system. With AutoML, people no longer need to design complex deep learning networks. In January 2018, Google launched the Cloud AutoML service, which released its own AutoML technology through the cloud platform. Even if you don't understand machine learning, you can train a customized machine learning model.   ◦ TensorFlow.js: TensorFlow.js was officially released at the TensorFlow Developer Summit 2018 at the end of March 2018. This is a machine learning framework for JavaScript developers. It can define and train models entirely in the browser. It can also import offline training TensorFlow and Keras models for prediction, and seamless support for WebGL. Use TensorFlow.js in your browser to expand your application scenarios, including interactive machine learning and all data stored on the client.   2.4 Reinforcement learning   Intensive learning seems to have a long way to go. At present, there is still no real breakthrough in the field of intensive learning. The study of intensive learning relies heavily on mathematics and has not yet formed a real industry application. I hope that more actual use cases of RL can be seen in 2019. This is a direction we need to pay attention to.   Google's new framework for reinforcement learning is Dopamine, the enhanced learning open source framework released by Google in August this year, based on TensorFlow. The new framework is designed with a clear and concise concept, so the code is relatively compact, about 15 Python files, based on the Arcade Learning Environment (ALE) benchmark, integrating DQN, C51, Rainbow agent Lite and Implicit on ICML 2018 Quantile Networks.   Third, algorithms and chips   The application of AI face algorithm and vehicle algorithm in the field of video surveillance is also becoming mature in 2018, and the accuracy rate is obviously improved. There is no big problem in commercial use, and the price seems to be reduced to an acceptable level. We mainly look at the progress of the AI chip.   At present, there are many companies that can design and manufacture AI chips in China: Huawei HiSilicon, Zhongxingwei, Cambrian, Violet, Horizon, Bittland, etc., but in this article we only discuss two: Bit Continental and Horizon.   3.1 Bit Continental   Bitcoin, which is famous for its mining machines and bitcoin, has officially entered the security market. It has also invested in the AI company Qianshitong. In Beijing, Anbo will bring their latest AI chips BM1880, BM1682 and counting intelligent server SA3. In just five years since its establishment, Bitcoin has developed into one of the most profitable companies in AI companies. With this alone, few companies can look forward to it. Bitcoin was established on October 28, 2013. At the end of 2015, the research and development of artificial intelligence chips began . In the first half of 2017, the artificial intelligence chip BM1680 was launched. In about one and a half years, this strength should not be underestimated. BM1680 is a dedicated deep-learning computing acceleration chip for the cloud. It is the first public chip to be released for this application. (We have heard many world firsts, no doubt this first is more valuable and also conforms to the big policy. And technology direction). In March 2018, the second generation of artificial intelligence chip BM1682 came out, which was about 5 times better than the first generation. Mainly used in security monitoring, data centers, super computing, robots and other fields. During the Beijing Anbo Conference, Bitcoin presented the two chips and the latest BM1880 chip.   Bitcoin believes that in the field of deep learning, especially in the field of reasoning, although the GPU has improved compared with the CPU, it still cannot meet the requirements of the deep learning algorithm for increasing computing power and power consumption. Therefore, the ASIC chip specially designed for deep learning needs to be constructed. Deep learning and optimized customization, TPU suitable for tensor calculation is the future. TPU is more suitable for tensor calculation and neural network. Compared with GPU, TPU has improved energy efficiency ratio by more than 10 times. TPU is designed for AI, with extremely fast speed, low energy consumption, low price and high cost performance. BM1682 is the second generation artificial intelligence chip launched by Bitland after BM1680, which is suitable for the reasoning of neural network models such as CNN/RNN/DNN. Compared with the BM1680, the BM1682 focuses on deep learning reasoning, with an average power consumption of 30W, and the actual performance is also more than 5 times higher than the BM1680.   The BM1880 TPU can provide 1TOPS computing power @INT8, which provides up to 2TOPS computing power under Wingoad convolution acceleration calculation. The BM1880 chip successfully streamed in July 2018. It is a deep learning reasoning artificial intelligence chip focusing on edge applications. It can provide 1TOPS computing power for 8-bit integer arithmetic and up to 2TOPS@INT8 under Winograd convolution acceleration. The specially designed TPU scheduling engine effectively provides extremely high bandwidth data streams for all tensor processor cores. The chip contains 2MB of memory and provides the best programming flexibility for performance optimization and data reuse. At the same time, BM1880 also provides users with powerful deep learning model compiler and software SDK development kit. Caffe, Tensorflow and other mainstream deep learning frameworks can be easily transplanted to BM1880 platform. Common neural network models such as CNN/RNN/DNN also Can run. The BM1880 chip can be used as a coprocessor for deep learning reasoning acceleration, or as a host processor to receive video streams, pictures or other data from an Ethernet interface or a USB interface, performing inference and other computer vision tasks; other hosts can also send video streams. Or picture data to BM1880, BM1880 to infer and return the result to the host.   3.2 Horizon   The Horizon is another company outside the Bit Continental that is focusing on security AI chips. Originally thought that the horizon is a chip design company, did not expect to bring a lot of security products and solutions in Beijing Anbo. Horizon Robotics is committed to being the global leader in embedded artificial intelligence platforms, empowering everything to make everyone's lives safer and better. Horizon is based on self-developed artificial intelligence chips and algorithm software, with intelligent driving, smart city and smart retail as the main application scenarios, providing customers with open software and hardware platforms and application solutions.   During the Beijing Anbo Conference, the horizon brought two embedded artificial intelligence vision processors. According to the official description, based on the innovative artificial intelligence special processor architecture BPU, Horizon independently developed China's first high-performance, low-power, low-latency embedded artificial intelligence vision processor for the Journey series of intelligent driving. Processor and Sunrise series processor for smart cameras. The first generation of processors was developed based on the Gaussian architecture and provided a complete solution of "Algorithm + Chip + Cloud".   The development line of the horizon BPU chip has gone through the Gaussian architecture, the Bernoulli architecture and the Bayesian architecture. We focus on the Rising Sun processor and target the smart camera. The Sunrise 1.0 chip was officially released and released in December 2017. Together with the journey 1.0, it became the earliest artificial intelligence chip in China to achieve mass production. The Rising Sun 1.0 chip integrates a deep learning algorithm with the ability to implement large-scale face detection tracking and video structuring at the front end. Currently, the Rising Sun 2.0 architecture has been applied to the Horizon XForce edge computing platform.   Fourth, it was originally IoT, which became AIOT in 2018.   The last air outlet before the AI is the IoT Internet of Things. IoT pays attention to the Internet of Everything. Any device with analog signals and digital signals can be connected, including various temperature, humidity, illumination, air quality and water level. Detection, gas, flow sensor, and through the digital switch signal, analog signal can control a variety of actuators, such as switch valves, switch lights, flow control, and so on.   The Xiaomi IoT platform has more than 85 million connected devices and more than 10 million daily devices. This is the number disclosed by Xiaomi founder Lei Jun at the first Xiaomi IoT Developer Conference at the end of 2017. He said that Xiaomi has now settled in the world's largest intelligent hardware IoT platform. Today, the millet IoT is connected to more than 100 million devices. According to forecasts, the economic value of the global IoT market is expected to reach $11 trillion by 2025, and the hardware market will reach $53 billion by 2020. Obviously this is a gold mine.   But these are not the key points. Although IoT can access voice equipment and video surveillance equipment, it is just a simple interaction and control before the emergence of AI, or even a function of viewing video and real-time monitoring. After 2018, AI and IoT finally showed signs of convergence. The first wave was the new revolution of smart speakers for smart homes. The smart speakers including Amazon, Apple, Baidu, Ali, Xiaomi, etc. can realize the voice of many IoT devices. Control, including various switch devices, RF control, infrared control devices; secondly, face device technology is also widely used in the IoT Internet of Things, through face recognition can determine the identity of people to set permissions, can be grading children Control, family and work environment can be deployed and controlled. These are some of the primary applications we have seen. I believe that there will be deep integration applications in 2019, and finally form the pattern of AIOT, which is AI Empowering IoT.   AI brings strong winds to traditional intelligent hardware and IoT-sensing devices, which has brought new life to the Internet of Things, which has been gradually diluted, and this shareholder's wind has blown from smart homes to wearable devices and VR. /AR extension. It is expected that in the next 1-2 years, AI will bring innovations in equipment interconnection, interaction, voice recognition, visual recognition, and interconnection. Deeply digging users' needs will form a new blue ocean, and it will also open artificial intelligence at the application level. More infinite possibilities, according to the latest reports, even the smart hotel based on the full control of AIOT has appeared, the core control device is a smart speaker.   Five, AI+ video surveillance commercial year   2016 can be said to be the first year of AI's large-scale commercialization. After two years of technological development and commercial landing, we can see the whole process of AI from research, papers, pilots to large-scale landing. At present, AI has two main capabilities: computer vision (the ability to see) and speech recognition (the ability to hear), which is the two main ways people perceive the world. Speech recognition is not within the scope of this article. We mainly discuss the ability to see.   If you counted from the first electronic tube camera developed by Panasonic in 1957, video surveillance has a history of about 61 years. It has experienced the analog era (1957~2004), the digital age (2004~2017), and the intelligent age (2017~). And the data age (2018~), before the intelligent era, the main functions of video surveillance were limited to three functions of monitoring, recording and playback. If you need to use video for auxiliary work, manual review is required, which is time-consuming and laborious, and the efficiency is not high because Large capacity usually saves the video for less than 90 days.   The emergence of AI technology completely solves the problems that traditional surveillance can't see and understand, especially when it comes to understanding functions and the human eye to understand an image. People all over the world do not need special study or training, and natural instinct can understand. An image, and the computer is missing. "Understand" means that the computer can read a frame of frames and continuous video images, which depends on computer vision technology (CV). After years of development, CV technology has begun to have full commercial capability in 2018, and has made great progress in face recognition, human body recognition, license plate recognition, and vehicle identification.   The application of face recognition can be divided into three categories: cooperative face recognition, semi-cooperative face recognition and non-cooperative face recognition. The first two types of recognition have always had a relatively high recognition rate, but no cooperative face recognition ( Dynamic Face Recognition) The true recognition rate is increased to a commercially viable stage (recognition rate over 70%) should be 2018, which we call the commercial year of AI+ video surveillance. The main applications of face recognition are face detection, face facial features, 1:1 face recognition, 1:N face recognition, and M:N dynamic control. The most widely used face recognition is 1:1 face identity check.   Vehicle identification is one of the earliest scenes of AI technology. It is more mature than human face, and license plate recognition is OCR recognition in a sense. At present, the commercial vehicle information structure processing mainly includes: motor vehicles (supporting 200+ vehicle brand recognition, supporting 4000+ sub-models and annual identification, supporting 7 categories of vehicle category identification, supporting 10+ body color recognition, and support Annual inspection, sun visor, seat belt, decoration, hanging, sunroof identification, driver face detection, front passenger face detection) and non-motor vehicle (based on ordinary video analysis, vehicle detection (two/three-wheeler), two rounds Car type identification, vehicle headlight shape analysis, body color recognition, fender detection and color recognition, tail box detection and color recognition, rear end advertisement detection, parasol detection and color recognition, rider helmet detection and color recognition , basic characteristics of people, characteristics of people's clothing).   Sixth, build a new ecological environment   In the market for AI-enhanced video surveillance, there are indeed a few unicorns, but there has not been a giant company that can take the world. Chips, algorithms, front-end devices, back-end devices, storage devices, platforms, systems, the entire video surveillance system is very complex, even Huawei can not provide all the equipment needed for video surveillance, so more companies choose to build AI security New ecology. The algorithm company assigns the algorithm to the chip, empowers the hardware, the chip is empowered to the camera, and the device is enabled to the structured host. The front-end hardware devices are used in conjunction with the software platform, which is almost as dependent on each other and compete with each other. Some companies expand their own sectors through acquisitions, mergers and acquisitions, and some companies build their own ecology through alliances and partners. All in all, 2018 is the year of ecology, chosen by many companies, and its importance is no less than AI empowerment.   Seven, ten million items per week   The author has collected about 2,339 project information since January 2018, including smart security related to video surveillance (Safe City, Xueliang Project, Video Access Control, Public Security Big Data) and Smart Transportation (Bayonet, Electronic Police, Urban Brain) Project 1936 The bidding amount is about 58.4 billion yuan, accounting for 9.7% of the total security market of 600 billion, close to 10%. Through the data analysis of these 1936 projects, the author found that there were three hot markets in 2018: Xueliang Engineering, Smart New Police and Zhi'an Community, and this was also revealed in the Beijing Anbo Fair in 2018.   7.1 Smart New Police   License plate recognition was first widely used in public security construction bayonet system and electronic police system. Face recognition followed closely, with scenes and needs that naturally landed in the public security market, identity matching, face recognition, big data comparison, Key personnel control has a natural advantage.   Before the AI did not empower the police, the video surveillance system was like a gold mine that was not excavated. The data was sleeping on the hard disk, and more than 90% of the data was covered by subsequent recordings. After AI, the system can automatically generate structured information (save pictures instead of saving videos), greatly reducing the pressure of video recording, and the largest application of structured video data is to achieve instant warning, quick detection of cases and crime. The core management element of the city is people, and the intelligent video surveillance system can identify a person well and generate tracks. If combined with multi-data collection, it can collide with more applications, such as ID card, MAC address, mobile phone number. Wait.   The AI's empowerment of policing has given birth to the construction of "smart new policing", and finally can use the video system to build a city's video big data project. Taking Guangdong's smart new police construction as an example, the overall plan is the “13847” strategy, which is a grand and beautiful vision, three steps to achieve the vision goal, eight innovative police applications, four wisdom empowerment projects, and seven-star plan. In Guangdong's planning, the most relevant to video surveillance is the video cloud enabling project, which establishes a video image resource service system based on the province's video images and multiple resource associations. The video cloud realizes the collection, aggregation, analysis, correlation and integration of video, portrait, vehicle, access control and other information, providing a resource sharing, capacity opening, security controllable and practical support for each police, city and grassroots combat departments. An integrated video cloud big data platform. From the perspective of video cloud empowerment project, its true technical core belongs to AI.   7.2 Xueliang project sings the protagonist   The large-scale application of video surveillance in the city belongs to the construction of Safe City (also known as Skynet Project). If it has been built for 13 years from the "3111 pilot project" in 2005, it will gradually mature in the region and build a safe city. Video surveillance usually covers large and medium-sized cities, and the smaller scale covers the county. The coverage of the county-level or urban-rural junctions is not much in the construction of the safe city. Therefore, the Xueliang project was born. The Xueliang project belongs to the “mass security prevention and control project”, which comes from “the eyes of the masses are sharp” in the safe city. In the case of saturation, the snow project in 2018 has gradually become the core of urban video surveillance construction. In 2015, Pingyi County was identified as the pilot county of Xueliang Project. In October 2016, the National Social Security Comprehensive Management Innovation Work Conference was fully deployed. “Snow Project”, after 2017 planning, was fully implemented in 2018. Among the 1936 projects counted by the author, 300 of them were related to Xueliang Project, with a bidding amount of 11.3 billion yuan, accounting for 19%.   7.3 Smart communities, Zhi'an community, and peaceful communities emerge in an endless stream   If the snow-light project is covered to the smallest unit and belongs to the community in the city, if you want to do the grassroots management work and video surveillance application, the best way to achieve it is to build a smart community. The community belongs to a relatively small scope and is easy to form. It is enclosed and can be managed nearby. It is a relatively easy-to-manage management unit. If you can manage each community well, you can manage the city if you expand it to the city.   If the video surveillance and smart communities are closely integrated, a new type of community will be created, namely the Zhian community and the safe community. Before the video surveillance was not AI, the potential of the video was not fully exploited. With the AI technology, the AI capable camera can realize face collection and license plate collection, which greatly improved the management level and governance level of the community. During the 2018 Beijing Anbo Conference, among the 105 panels (incomplete statistics) of the “Public Security Technology Innovation Application Results Exhibition”, 15 panels showcased the solutions or achievements of the Zhi’an community, which is second only to the Xueliang project. The Zhian community will be a new direction for future police construction.   Eight, AI morality   Although AI has given us a lot of new technologies and applications after the city, it also brings some problems. These problems are more or less presented in 2018, that is, the ethical issues of AI. AI abuses were frequently reported in 2018: Facebook AI helped Trump be elected president of the United States, Goggle and the US military teamed up to develop AI weapons, and Microsoft provided cloud computing and face recognition services to immigration and customs enforcement (ICE) These stories have led to the discussion of the industry's extensive AI code of ethics**, and even some companies have disclosed the company's AI guidelines (such as Microsoft has proposed six: fairness, transparency, accountability, non-discrimination, informed Agree, legal surveillance). However, some people believe that AI morality is still a gray area, and there is no unified framework yet. It is expected that more regulations will be introduced in 2019.   There are two things related to AI ethics in 2018. The European Union launched the Global Data Protection Regulations (GDPR), which aims to improve the fairness and transparency of the use of personal data. The Regulation gives individuals the right to control their personal data and understand how personal data is used. This was followed by the “Cambridge Analysis” scandal, which cast a shadow over the entire data science community, which triggered a deeper level of ethical discussion.   Outlook 2019   Predicting the future development trend is actually a very difficult thing. From the perspective of personal understanding, the author hopes to make breakthroughs in the following aspects in 2019, including the ReID pedestrian recognition system, LoRa long-distance communication technology, chips, A new model of smart city construction, brain engineering, data age, three-dimensional fusion, voiceprint applications, etc.   First, ReID   The next hot spot of video surveillance in 2019 will appear, and ReID is definitely one of the hot spots. Face recognition technology is good, but what if you can't see the face? This is a question before us. In fact, in the cameras inside the city, in addition to the camera camera and the electronic police used to capture the camera, the camera that meets the face collection standard is relatively rare, accounting for a small proportion. To maximize the potential of security cameras in traditional safe cities and Skynet projects, ReID is the best choice.   ReID (Person Re-identification), also known as pedestrian re-identification, pedestrian re-recognition, cross-mirror tracking, is a technique that uses computer vision technology to determine whether a specific pedestrian exists in an image or video sequence. It is widely regarded as an image retrieval sub- The problem is currently mainly used in the security field. The combination of future and face recognition can be applied to more and more rich scenes.   ReID has not been noticed by many people in the past and has been using the product for a very long time. Since there was a technology that relied entirely on traditional computer vision or machine learning in the early days, there was basically no significant breakthrough. ReID itself is a very difficult problem, it is to identify the same person from different videos. Different lighting conditions, different resolutions and angles of the region of interest, general occlusion of the target, people wearing similar clothes, etc., can cause difficulty in identification. For the monitoring field, ReID is extended in practical applications, that is, it is desirable to associate objects in different videos, and to find out the objects through effective methods. This can be done in such a way that the entire monitoring operation is complete and can be seen. To the big picture. Jiadu Technology's shareholding in the second half of 2018 made a major breakthrough in ReID. The first hit rate (Rank-1 Accuracy) on the Market 1501 has reached 97.1%, surpassing the human eye recognition ability (94%). And refreshed the 96.6% world record announced in April this year, we have seen the broad market prospects of ReID.   Second, LoRa technology   LoRa is the wireless standard for ultra-excited communications from Semtech. Ganquan, director of market strategy at Semtech, pointed out that Lora's market growth is very fast, with global growth of around 50% per year, and China will be even higher. LoRa's applications in smart cities include energy management, smart building, smart production, and smart agriculture.   So why is LoRa suitable for smart cities? Gan Quan said: "The most important thing for IoT is to have more connections, successfully retrieve the data, and let users get a better experience in the control and use of the connection. The best way is to use low-power LoRa technology. "LoRa is a low-power WAN, derived from the two words Long Range. How to understand the super long distance of LoRa? Gan Quan pointed out that there are already satellites equipped with LoRa technology. The satellites orbit the earth and are 600-1600 kilometers away from the Earth. The equipment on the earth can transmit data with the satellites. Being able to use satellite communications is enough to prove the power and advancement of technology and the various possible application prospects. However, LoRa does not transmit large data streams such as video. It can only transmit simple data such as control switch signals or brightness, table readings, etc.   In addition to the outstanding features of ultra-long distance, LoRa is also the lowest power communication technology in long-distance communication, which is also very important, and only requires one gateway to manage a very large number of devices, which also reduces LoRa on the one hand. Deployment costs. This is very important in the application of IoT technology in the city. The three points of long distance, low power consumption and large scale are enough to kill. The author believes that LoRa technology can achieve deep integration with AI and video surveillance, especially in the application of video surveillance trigger linkage alarm.   Third, the chip is faster and more complicated   Limited by the material mastered by the author, in December 2018, Zhang Yongqian, general manager of Horizon Smart Chip Solutions Division, gave a speech in Shenzhen on the "Horizon AI Chip Solution" to see the development of the chip industry in 2019. I think the horizon is a The typical representative of the enterprise.   Horizon believes that the chip brings the era of change, that is, from the era of model innovation to the era of technological innovation, that is, the trend of 2019, emphasizing "technology to improve social productivity and improve efficiency." The evolution of AI chips from CPU, GPU, FPGA, TPU to BPU, one level one step, the top of the pyramid is TPU and BPU (may not be recognized by some readers). The biggest change that the Horizon made in 2019 was probably to package the "algorithm + chip" to provide customers with a one-stop solution, without the need for expensive algorithmic integration, which may be good news for SI.   In 2018, the horizon has released the processor of the Rising Sun 1.0 series. According to the latest plan, they will launch the Rising Sun 2.0 series processor in Q19 in 2019. Compared with the 1.0 series, there is a big improvement. For example, the Rising Sun 2.0 series supports 1080P@30fps. ×4 Camera (1.0 is a 1-way camera), front-end & edge products (1.0 only supports front end), large storage face recognition (500,000 people dynamic comparison), pixel-level semantic/action behavior analysis (previously not), Multi-channel video structuring (human/car/non-motor vehicle, etc.), comprehensive upgrade open training platform, based on Bernoulli architecture. The most powerful is the introduction of the Bayesian 3.0 series based on the Bayesian architecture in the first quarter of 2020. The speed of the chip launch in 12 months has already broken Moore's Law. The new Rising Sun 3.0 series will support 4K@30fps, front end & edge according to the plan. Product & intelligent server, support Monte Carlo decision search, support complex network structure such as RNN, simultaneously handle 12-channel video, semantic 3D environment modeling, dynamic path planning. From the planning of the horizon, we can see the general trend of chip development in 2019. The development of the horizon is the most optimistic of the author.   It is worthwhile to integrate the detection of existing objects in the form of detection frames for object detection and recognition. In 2019, it will evolve into pixel-based semantic segmentation and understanding, that is, three-dimensional object detection analysis, which can be based on people. The contours and the contours of the car are identified, which is an innovative technology. A single chip can realize large storage capacity deployment and achieve dynamic comparison of 500,000 people. That is to say, if you control 2 million key personnel, you only need 4 camera coverage to implement front-end control. This will definitely have a great application. prospect.   Based on the Rising Sun 2.0 series of solutions, the hardware supports video structure processing, which can realize multi-target object detection and tracking of motor vehicles, pedestrians and non-machines at the front end, which may bring pressure to software video structuring manufacturers, but this is 2019 Trends; behavior analysis, multi-camera fusion ReID solutions will also appear in the products of the Rising Sun 2.0 series.   In 2019, the chip will be faster, more complex, and more powerful.   Fourth, a new model of smart city construction   Since the introduction of IBM, the smart city has been widely used and applied in China, and has achieved many results and summarized many experiences. The traditional smart city construction mainly focuses on social security, industry efficiency and people's livelihood services, which greatly improves the efficiency of urban governance.   However, with the development of AI technology in the last two years, the ability of AI to empower cities has gradually emerged, not only in video surveillance, but also in the community, in medical care, in education, in the financial industry, many new applications have been spawned. Taking Huawei as an example, at the 2018 Huawei Total Connection Conference, Huawei Cloud launched the EI City Agent to provide better urban transportation, water, environmental protection, and gas solutions. This conference marked Huawei's opening of the combination of chip + frame + platform + service full stack synergy, and began to fully benchmark the international AI giant.   In addition to Huawei, Tencent, Ali, and Baidu have also launched new smart city solutions. Older integrators like Jiadu Technology have also launched a new smart city construction program (including the urban brain). We believe that in 2019, there will be a new type of new smart city construction model with multiple types and multiple samples, and this will be thanks to AI, and AI will be everywhere in the city.   Fifth, brain engineering   In the end, AI still simulates the human brain. If we put the application ability of AI empowerment in the city, it is to build the city brain. The idea of the urban brain is more experienced than the smart city, the central nervous system of urban management, and the concept may be smaller than the smart city. If the urban brain is broken down, there will be a city police brain and a city traffic brain.   In the past year, we have seen that the land is better including Ali's urban brain model, Fang Wei (Jiadu Technology's corporate brain model) and Baidu's urban brain model, but in terms of actual landing, the traffic brain is the most First landing, this is because the transportation infrastructure is better, traffic lights, signal lights, electronic police and bayonet cameras all over the city, which are closely related to video surveillance.   More noteworthy is the Baidu brain, similar to Google's brain, Baidu hopes to develop a more powerful brain system, empowering the entire city. At the Baidu AI Developer Conference in July 2018, Baidu Brain announced the upgrade to version 3.0, which first proposed "multi-modal depth semantic understanding" in the industry, forming a AI full stack from chip to deep learning framework, platform, and ecology. Technical layout, this is also the most complete and cutting-edge AI technology platform in China. Baidu Brain 3.0 also opened more than 130 advanced AI capabilities, continuing to equalize the empowerment of developers.   The author believes that in 2019, there should be no less than 10 large and medium-sized cities to start the construction work of the city brain.   Sixth, DT era of video big data and video cloud   The data age (DT) may have been raised for many years, but for the video surveillance industry, the real DT era began in 2018, and began to land in 2019.   As mentioned above, the development of video surveillance has gone through four eras. In 2018, it has fully entered the data age. Cloud computing and big data are no longer fashionable words, and they have been deeply penetrated into all aspects of social governance. After the unstructured video image data is structured, video image big data can be formed. The data can be divided into four categories:   ◦ Panorama data . Contains data such as people, cars, objects, mobile phones, access control, WIFI, IoT perception, maps, addresses, house numbers, grids, population, houses, units, and urban parts within the spatial dimension. The panoramic data reflects the full data and multi-dimensional data analysis in multiple scenes.   ◦ Full data . On the basis of the panoramic data, including the time dimension, full-time data, including data such as tracks, activities, events, and the like.   ◦ Global data. The association between the data built on the panoramic data belongs to multi-dimensional related information, which is collected by multiple channels, multiple perspectives and multiple sides. A model that contains all the information about the system, enabling data correlation, collisions, and multidimensional perception.   ◦ holographic data . The global data and the video image are merged to generate a stereoscopic space, multi-dimensional, and interrelated all-time-space data. Typical applications include 3D holographic projection, virtual display VR, and enhanced display AR. Holographic data reflects social attributes and reflects the value of data.   The author's judgment, the characteristics of video surveillance in the data age is that it can be comprehensively viewed, automatically viewed, and associated.   ◦ overall look . Video images are integrated and shared across the network. Across-system, cross-regional sharing of multidimensional data across a wide range.   ◦ Automatically watch . High-density, high-computing, multi-algorithm framework, and hundreds of billions of pictures are searched in seconds, which is fast and accurate.   ◦ of relevance . Collision analysis of video big data and resources such as social, network, government affairs, and police big data. Realize "image event correlation", "face, vehicle, mobile phone and other multi-track integration" and other applications.   Seven, 3D, AR, VR deep integration application   In 2018, Beijing Anbo will be the vane of the development of video surveillance industry. We can see that the video application system has gradually transitioned to the deep integration of 3D, which is to deeply integrate 3D map, AR, VR three-degree technology and video and data, and then Developed a new application.   The basis of this deep fusion application will be video surveillance networking platform, video analysis platform, video image information database, and a city management basic information data platform (also known as a standard three real grid system), and these data Can be combined with 3D, AR, VR. For example, we can directly embed multi-dimensional data into a three-dimensional map, and directly embed the video into the map through the AR enhanced display method to realize visual real-time urban image presentation, and directly view various types of data through VR technology. In the eyes of the people, timely access to information data is achieved.   Eight, voiceprint + CV technology   The voice author in the audio and video system believes that it has not been fully tapped. We have seen that many CV head companies have made great investments in voice, and there are many voice recognition companies on the market, although in 2018 we I didn't see the application of a particularly good combination of voiceprint + CV technology, but I believe that the combination of the two will have great potential. Taking the access control system as an example, we can use the dual mode of face + voiceprint, input a face and then pick up a person's voiceprint for secondary confirmation, or enter a person's voiceprint and then call a person. By performing a second confirmation on the face, an accurate match of 1:1 can be achieved, which can greatly improve the accuracy and false positive rate of face recognition. I hope that in 2019, we can see new applications and new hot spots in the industry. 

  • facebook