A Design Pattern of IAPVS Platform Based on Distributed Edge Computing

Traditional video surveillance is difficult to effectively cope with real-time interaction between end side devices of over ten thousand scale and data centers, and these devices are widely dispersed and deployed on a large scale. They cannot meet the needs of diverse scenarios monitoring and smart applications such as security, urban comprehensive management, ports, mines, water conservancy, power, industrial manufacturing and other fields. Based on the analysis of the problems in the popular intelligent video surveillance system, this paper proposes the design mode of an intelligent video surveillance intelligent analysis platform (IAPVS) based on distributed edge computing by improving the existing centralized video cloud computing architecture.This design pattern provides a new overall design solution for the development of platform of intelligent video analysis suitable for diverse scenarios and smart application fields.


Introduction
Video surveillance technology has roughly experienced the development from the traditional analog video surveillance, digital video surveillance, network video surveillance, cloud-end mode of video surveillance, to a new generation of video surveillance technology based on edge computing.Video surveillance based on network video surveillance and cloud-end mode has been widely used now.At present, the demand for intelligent video surveillance is increasing in the fields of security, urban comprehensive governance, port, mining, power, industrial manufacturing and other fields is increasing, which has given birth to the application of a new generation of video edge computing technology [1][2][3].
With the increasing demand for intelligent video surveillance in diverse scenarios and intelligent application fields such as security, urban comprehensive management, ports, mines, water conservancy, electric power, industrial manufacturing, traditional video surveillance is difficult to effectively cope with the wide-area decentralized deployment of more than ten thousand orders of magnitude front-end acquisition equipment.Problems such as end-to-end intelligent perception, realtime interaction with data centers, massive data and task computing, and accurate and mature application of AI algorithms cannot meet the needs of diverse scenarios and smart applications.The current popular centralized video cloud computing architecture, namely intelligent video surveillance system based on cloud-end architecture, faces challenges such as transmission delay, network capacity, and computational processing efficiency.
On the basis of analyzing the development of Intelligent video surveillance technology, Using an approach different from the existing centralized video cloud computing architectures,this paper proposes a design pattern of Intelligent analysis platform for video surveillance (IAPVS) based on distributed edge computing, it provides a new overall design scheme for the development of video intelligent analysis platform, which is suitable for various scenarios and intelligent applications such as security, city comprehensive administration, ports, mines, Water Conservancy, electric power, industrial manufacturing and so on.The following contents are arranged: the video cloud computing architecture of distributed edge computing, the logical design of IAPVS platform, and the workflow process of IAPVS platform are introduced.2.Video cloud computing architecture for distributed edge computing.

Thelimitations of Centralized Video Cloud Computing Architecture
Intelligent video surveillance systems the so-called centralized video cloud computing architecture (cloud-to-cloud Architecture) , in which end-side devices connect directly to the cloud or data center, where decisions are made by learning, reasoning, and so on, decisions made may be sent to end-side devices for execution.The centralized video cloud computing architecture is illustrated in figure 1. Centralized (that is, centralized) cloud computing architecture for intelligent video surveillance systems, the camera on the device at the edge of the network transmits video directly to a networkcentric cloud server over a wireless network/private network and the Internet.The cloud server stores the video after it is acquired.The real-time processing and analysis of massive video streams requires the use of machine learning algorithms such as depth convolutional neural network, while the vast amount of visual analysis computation can only be run in real time on specialized hardware such as the graphics processing unit, it is necessary to offload the computing tasks to the devices with strong computing power for reasoning analysis, and then integrate the reasoning results in the cloud.In some scenarios, the cloud also returns the final result to the terminal device at the edge of the network [1][2].With the popularity of smart phones and the large-scale deployment of surveillance cameras, for wide-area, more than 10,000 end-to-side (and heterogeneous) devices connected to the data center, high-delay can not meet the low-delay transmission requirements; The huge amount of GB-level data collected by the front-end equipment is uploaded to the cloud, and the network bandwidth is under great pressure; therefore, the centralized video cloud computing architecture for video stream transmission processing, computing, storage and management, the ability to deal with transmission Learn Action Upload Original data Send Operation learn action delay, network capacity, and computational efficiency is severely limited, and can not meet the needs of large, real-time video stream intelligent analysis.

The Video Cloud Computing Architecture of Distributed Edge Computing Collaborates with "end-edge-network-cloud"
In 2014, the European Telecommunications Standards Association introduced the new concept of Edge Computing (EC) , which provides users with internet technology and cloud computing capabilities through wireless access networks located near mobile users.In 2017, Microsoft put forward the design of Edge Computing real-time video stream analysis system for intelligent traffic analysis.Since then, many edge computing intelligent video surveillance systems and research began to emerge, and gradually deployed to security, urban comprehensive management, ports, mines, Water Conservancy, electricity, industrial manufacturing and other fields of application.
Edge computing realizes data collection, analysis and processing at the data source by lowering the traditional cloud computing data center to the edge of the network, it not only guarantees the requirement of real-time processing, but also greatly optimizes the utilization ratio of network resources.Edge computing is designed to bring computing tasks down from a cloud server in the center of a network to an edge server or smart terminal device that is physically close to the video source, in an edge computing architecture, computing tasks can be unloaded into device-edge-cloud hierarchies [1][2][3].Smart devices and edge servers with some computing power can handle most of the storage and analysis tasks directly near the video source.Cloud servers provide computing support and data fusion only when necessary.Therefore, edge-based video stream analysis system will significantly reduce the bandwidth of the Internet upload.In addition, the edge server is physically close to the video source, and its round trip time (RTT) is almost negligible compared to the cloud server.Edge-based video stream processing system only needs to consider the transmission delay of video stream and the delay of video analysis.At the same time, compared with the Internet, video source and network edge connection is relatively simple and controllable, to ensure adequate transmission bandwidth and stability.Cloud privacy leaks can also be effectively avoided by processing video data directly at the edge or uploading it after desensitization at the Edge [1,[4][5][6][7].
In recent years, the development of edge computing technology to the intelligent video monitoring system to the front, at the scene of the surveillance / near field nearby deployment of several video edge computing gateway or "edge to the edge cloud" edge of the micro server, no longer need to all the original data to the cloud to process, and then send back, but directly in the edge side can be completed processing, calculation more close to the data source, response processing delay is lower.The edge side preprocesses the data, filters the low-value data, reduces the transmission with the cloud, and reduces the network bandwidth pressure.At present, the edge computing gateway and the edge side microserver have been put into the market, and the performance has also been greatly improved, which provides an important support for the intelligent video surveillance system using the video cloud computing architecture of edge computing.Mainstream edge computing vendors and their devices include: the Jetson platform launched by NVIDIA Corporation and computing that can efficiently use GPU to achieve a variety of deep learning.;Atlas 500 intelligent station launched by HUAWEI China; Google designed deep learning accelerator Edge TPU enables high-definition video real-time deep learning reasoning with high energy consumption ratio; 5G industrial video edge gateway launched by Shanghai Guomao Digital Technology Co., Ltd.; 5G edge computing terminal product of Hangzhou Danghong Technology Co., Ltd.China; EasyCVR intelligent edge gateway of Anhui Xufan Information Technology Co., Ltd.China; AI terminal of China Lianbao (Hefei) Electronic Technology Co., Ltd.
The video cloud architecture for decentralized edge computing is shown in figure 2. The design of video cloud computing architecture for distributed edge computing breaks through the limitations of traditional video monitoring GPU server cluster deployment; video edge computing will include face detection, capture, and recognition, intelligent front-moving end-to-side cameras such as structured video, vehicle detection and recognition, end-to-side/Edge/center with software definition capability; supporting end-to-side, side-to-side, net-to-cloud collaboration, it supports on-demand deployment and dynamic loading of corresponding end-side/edge/center AI algorithms to form an efficient intelligent supply integrated with the whole network, and enhances its fast adaptability to the application of AI video analysis overloading, adapted for more flexible video intelligent deployment and diversified scenario applications [8][9][10].

Video Surveillance Intelligent Analysis Platform Based on Distributed Edge Computing
Important indicators to measure the performance of intelligent video streaming analysis include: latency, analysis accuracy, bandwidth, throughput, energy consumption, memory usage, etc. (see table 1).
Based on the important performance index of Video Stream Intelligent Analysis, the video monitoring intelligent analysis platform (IAPVS) for distributed edge computing is designed.IAPVS platform can container and resource manage all functional modules in the system, and container and resource manage all functional modules in the system, contains model repositories, query optimizers, and custom key-value storage modules.The model warehouse maintains the parameters, uses and resource occupancy of the deep learning model, and can choose the suitable model deployment according to the application requirements and resource status, it also supports the parameterization and calibration of the model.The custom key storage module acts as a public cache of the video analysis module, storing part of the mapping from the video frame input to the analysis results, reuse partial results for the same video stream when reasoning on different analysis nodes.The IAPVS platform also provides data storage and mining capabilities, including cross-camera and cross-history video analysis and further mining of sensor metadata, as well as analysis of system logs and performance data, and persist the data [1,9,10].

Indicators Explain
Delay End-to-end delay, response time, from video frame decoding, data transmission, model reasoning, to the final analysis of the total time.

Video analysis accuracy
A measure of the predictive accuracy of video analysis algorithms and models.Different visual tasks often correspond to different accuracy indicators.

Tape width
The amount of data per second transferred between the terminal device and the edge server, the edge server, and the cloud server.

Throughput
The number of devices that can be simultaneously serviced by the Edge real-time video analysis system per resource, or the number of analysis requests that can be simultaneously serviced by the same system per unit time.

Energy consumption
The energy consumption index is of great significance to the battery-based mobile terminal equipment.Video Monitoring Intelligent Analysis Platform of each module, in unit time or a single analysis of the energy consumption task.

Memory occupation
The amount of memory the computer vision model occupies on the device.Because deep learning models usually have tens or even hundreds of millions of parameters, it needs enough memory space to load.
The IAPVS platform includes two core parts: video edge computing nodes and data center.The logical design block diagram of the IAPVS platform is shown in figure 3.

Video Edge Computing Node
Edge computing gateways (or edge cloudlet servers) are deployed in the field/near field to connect sensors/end-side devices.Through edge computing and the deployed learning model, the raw data are filtered to generate high-value important data and then uploaded to the cloud and data center.
In figure 3, for the gateway of the video edge computing node, the main module design adopts Chinese domestic Feiteng embedded CPU, ARM + Cortex multi-core hardware architecture and Linux software architecture, with the maximum onboard board of 4GB LPDDR4 SDRAM of memory.The digital video coding standard adopts Chinese standard AVS +.The gateway supports real-time transmission, storage and processing of high-definition video streams, supports access protocols for connecting various devices on the gateway side, and supports the security and trusted management of the full link communication between the gateway side and the video main control center.The gateway has the AI computing power of 21 TOPS floating point operation for deep learning ability.

Data Center
The data center includes video big data platform, distributed machine learning platform, video data analysis computing platform and model warehouse.A large number of scene models such as face recognition, wearing, foreign object, equipment status, and optical character recognition (OCR) are loaded and accumulated in the model warehouse.When users have new requirements, the models can be quickly corrected and reused.The video big data platform has preset the original training data, which can train the computer vision (CV) model for the distributed machine learning platform, package the trained model to the shelf, and launch the model online after the video data analysis and computing platform.It can also directly provide application programming interface (API) services for users.After the statistical model and the CV model of learning and reasoning are online, the accuracy will slowly decrease as time goes by, so the model needs to be updated in time.The video data analysis and computing platform can combine the PipeLine modeling workflow with the new sample set and the predicted value of the API, do the iteration of the model, and realize the new model to cover the old model.[11][12][13][14]

Conclusion
Based on the analysis of the application requirements of intelligent video surveillance and the analysis of the limitations of the so-called centralized video cloud computing architecture (cloud-end architecture), the paper adopts the video cloud computing architecture of decentralized edge computing, and proposes a design pattern of intelligent video surveillance analysis platform based on decentralized edge computing (IAPVS).The logic design and workflow process design of IAPVS platform are given, which provides a new overall solution for the development of intelligent video analysis platform suitable for diverse scenes and field intelligent applications.

Figure 1 .
Figure 1.Centralized video cloud computing architecture

Figure 2 .
Figure 2. Video cloud computing architecture based on decentralized edge computing

Figure 3 .
Figure 3. Logic design block diagram for IAPVS platform The video surveillance intelligent analysis platform (IAPVS) of distributed edge computing is designed based on the video cloud computing architecture of distributed edge computing.The IAPVS platform consists of two core components: the video edge compute node and the data center.IAPVS platform logic design block diagram is shown in figure 3.