, GPU is mainly used in data center. The disadvantage is high power consumption. In security applications, GPU chips are basically monopolized by NVIDIA. GPU chips in the security industry are monopolized by NVIDIA. Compared with GPU, FPGA central reasoning and data center are also widely used. FPGA has obvious advantages in power consumption. In security applications, the main FPGA manufacturers include xilinxintel, original Altera, etc. Due to the diversity and complexity of end-to-end applications and the requirement of high cost performance, ASIC is mainly used in end-to-end reasoning.
There are many ASIC manufacturers, such as Cambrian, Hisilicon, horizon, bitcontinent, etc., and many solutions are provided at the same time. ASIC competition in the security market in 2018 is very fierce,. Among them, the layout of Hisilicon is very intensive, and deep learning is the core of artificial intelligence. Software framework technology is still in the hands of technology giants such as Amazon, Microsoft, Google and Baidu. It provides an integrated software toolkit for application development, and the algorithm framework is the key link in the establishment of artificial intelligence core ecosystem. Realize the modular encapsulation of the algorithm.
It includes various applications and algorithm toolkits developed for algorithm implementation, and provides algorithm call interface and other services for upper application development. Also began AI algorithm technology layout. With the further investment of enterprises, the algorithm enterprises in the security industry can be divided into two categories. The first category is Shangtang, Kuangshi, Yuncong, Yitu, Zhongke detective and other CV enterprises; The second category is Haikang, Dahua and Yushi. The accuracy of video recognition algorithm has been greatly improved.
For example, face recognition algorithm has reached a high level under specific conditions. Image classification, object detection, etc. The recognition rate of computer has far exceeded the average level of human beings. 4. Product and industry application CV manufacturers and cloud platform suppliers also gradually began to provide products and industry applications. In addition to the conservative security manufacturers such as Haikang, Dahua, Yushi, Keda, Tiandi Weiye and Dongfang Wangli. Enterprises have basically completed the serialization of edge intelligent products. The parking lot management system is enriched with edge and center products, except for the cloud center products mentioned above.
Users began to put forward higher requirements for application services. From the current industry situation, whether it is radical security enterprises or CV and cloud platform enterprises, although there is some improvement in business applications, it is still dominated by typical general applications. Users pay more attention to practical application. At present, there has been a new trend of "no AI and no security" in the security industry. With the penetration of AI security industry and the research and development of deep-seated application technology.
At present, it is an indisputable fact that all security monitoring manufacturers have a full line of product AI, and it has also become a new strategy for all manufacturers. With the in-depth implementation of AI security industry, the product form and application mode in the field of AI security, especially in the field of video surveillance, have also begun to stabilize. AI technology in the security industry mainly focuses on face recognition, vehicle recognition, pedestrian recognition, behavior recognition, structural analysis, large-scale video retrieval and so on. The former refers to application scenarios with controllable conditions such as light and angle. AI application scenarios in the security industry are divided into bayonet scenarios and non bayonet scenarios.
Mainly vehicle bayonet and face bayonet; The latter refers to the general security monitoring video scene. Among them, the bayonet scene accounts for about 1% - 3% of the total number of surveillance cameras, and the rest are non bayonet scenes. Surveillance video 1 bayonet scene: the key points of face identification application are equipped with face capture cameras, and the face identification application is represented by the control of personnel in the public security industry. The captured faces are analyzed and recognized by the back-end face recognition server, and compared with the face blacklist database. With the enhancement of personnel control application, the effect has begun to show.
For example, the recent "Jacky Cheung concert" captured the suspect is the identity confirmed by the bayonet scene. 2 bayonet scenario: face authentication applications such as face access control, face quick access door, face attendance, personnel identification, etc. face authentication applications are becoming more and more popular. Rare face whitelist applications have been implemented in many industries. It is widely used in enterprises, various parks and other scenarios. In addition to realizing the basic face recognition application, face access control can also prevent face counterfeiting through photos, videos and other behaviors, and effectively ensure the safe control of entrance and exit personnel and daily personnel management. 3 bayonet scene: vehicle recognition application license plate recognition makes "looking for people by car" a reality. Vehicle recognition technology is one of the most mature and effective technologies in public security practice.
With the help of popularizing vehicle checkpoints on major traffic roads all over the country. Shengli helped the police solve various cases. Vehicle recognition technology has developed from the primary vehicle recognition application stage based on license plate to the accurate vehicle recognition application stage such as vehicle type recognition and license plate recognition. 4 non bayonet scene: Video structured analysis and rapid retrieval application the video structured business function is to classify and detect motor vehicles, non motor vehicles, pedestrians and other moving targets in the video; At the same time, the small target image and the large scene image are extracted and written into the storage device for video structured analysis and rapid retrieval applications. It is convenient for subsequent rapid query and intelligent retrieval. Through the rapid analysis and extraction of the characteristic attribute information of the target of interest in the video through the video structured service, users can efficiently obtain the clues related to the case and events, and promote the video data from seeing to understanding in the era of great security. 5 non bayonet scenario: the application of behavior analysis assisted security can be applied to the machine identification of key area prevention, important goods monitoring, suspicious dangerous goods left behind and other behaviors; It can also alarm people's abnormal behavior, and behavior analysis can assist security application. Analyze and handle the abnormal behavior of personnel through the behavior analysis system.
It greatly improves the application efficiency of video surveillance. Eight restrictive factors for the scale application of intelligent security, but in the production and practical application, although the artificial intelligence technology has developed rapidly. There are still many problems. In the past few years, the popularity of artificial intelligence is very high, but in fact, only the "conceptual model" has been established, and the ideal effect of "effective utilization" has not been achieved. At this stage, there are eight main factors limiting the scale of application: high interest, algorithm scenario restrictions, high distribution difficulties, higher network and security requirements, lack of systematic top-level design for in-depth application, lack of industry norms and evaluation system and user learning and organization to ensure higher interest impact "There are many factors for the large-scale application of Security AI product solutions, and the interest is high. At present, the high interest is one of the many reasons.
From the cost proportion of each part of a typical medium and large-scale rural public security video surveillance networking project, it can be clearly seen that the cost is an important bottleneck in the development of Security AI. It is also the main bottleneck at this stage. The algorithm is highly limited. Artificial Intelligence Computing The generalization ability of the method is a temporary problem faced by the pattern recognition problem. Therefore, in practical use, the neutral energy of the trained model is often significantly reduced when it is used in changing scenes. It is necessary to strictly define the scene, or position the intelligent algorithm as an auxiliary function insensitive to indicators. In childish applications, such as passing and violation in intelligent transportation Chapter capture and the comparison of human certificates at airport stations all require specific engineering device schemes.
This approach effectively realizes the commercial value under the condition of insufficient technology, but the disadvantages are also obvious: on the one hand, the transformation of existing equipment needs to increase the construction cost, which affects the penetration of artificial intelligence algorithm into traditional applications; on the other hand, it also limits the efficiency of obtaining effective materials Rate, which affects the further improvement of algorithm indicators. Only in specific scenes can we adhere to a good recognition rate. It is difficult to arrange cameras for face recognition. Artificial intelligence often has specific scene requirements. It is necessary to strictly comply with the policies and Specifications issued by the Ministry of public security, which greatly reduces the application space of face recognition and greatly improves the construction difficulty.
The number of accidents occurred According to more and more reports, the next generation of human-computer interaction technology is developing with the application of intelligent technology in the security system. The attendant problem is how to make users quickly understand data, that is, the problem of data visualization. The next generation of human-computer interaction in the security industry will develop towards stronger operability and three-dimensional, interaction towards stronger mutual motion, and application Further drive or assist user decision-making. Therefore, the system design and project practice ability of the parking lot management system are improved from time to time. The landing of intelligent business applications needs to be based on reasonable interest control, qualified construction quality, perfect data integration and supporting management mechanism.
Then, the algorithms and models of supporting scenarios are based on high An effective computing framework transforms data into visual user services. Intelligent business application is a systematic project, including architecture, algorithm, computing, data, application, engineering and management process. It is necessary to strengthen the ability of systematic top-level design from time to time and improve the project practice ability. China has a demand for about 50 million cameras a year, and the stock of non bayonet videos is gradually applied According to statistics, only about 500000 smart cameras are effectively used, accounting for only about 1%, while up to 99% of cameras cannot be endowed with "intelligence" Attribute. This means that security AI has just entered the primary stage.
The generalization ability of artificial intelligence algorithms in non bayonet scenarios is one of the main bottlenecks in the implementation of security. Compared with radical pattern recognition methods, the generalization ability of deep learning algorithms and the adaptability to complex scenarios have been significantly improved with the support of big data. The technical foundation and product integration of intelligent security have become mature , Zhidong thinks. Therefore, the proposition in the next stage is how to systematically arrange the scale. Challenges and opportunities coexist. From the continuous innovation of technical means to the naive landing of product form, intelligent security still faces many problems, such as high interest, difficult engineering layout, large limitations of algorithm scenarios, lack of in-depth application, lack of systematic top-level design and lack of practical satisfaction How to solve these problems is related to whether intelligent security products and solutions can really take root.