Article Preview
Top1. Introduction
The Internet of things connects physical devices and cyberspace through chips and sensors. In a large multi geographical sensor network (GSN) (da Silva 2018), a large number of useful data can be collected to improve people's cognitive ability of the environment. For example, as the main type of data collected from surveillance cameras, visual data has been widely used in automatic monitoring applications, such as road monitoring (Fedele 2018), pedestrian flow detection (Hoshino 2016) and abnormal activity warning (Dhiman 2019). With the development of image processing and data analysis technology, we can utilize the images from cameras to perform data analysis, such as quantity statistics, face recognition, license plate recognition, real-time tracking. The intelligent surveillance system across multiple cameras (Adrian 2018) is a typical application of camera images. The intelligent surveillance system across multiple cameras can realize intelligent data analysis and automatic decision-making. However, large-scale intelligent surveillance system with hundreds of surveillance cameras often suffers serious system performance degradation problem due to magnitude of data. It is a challenging task to deal with raw video data with several TB in time.
In order to process the data from massive Internet of Things (IoT) sensors, some researchers propose to store and analyze the data by using cloud computing (Hosseinian-Far 2018). However, in an intelligent surveillance system, realizing the data analysis in real time needs high computing power which may far exceed the computing power of ordinary cloud computing. Merely using cloud computing to achieve video processing will bring very expensive procurement cost for cloud server. Additionally, in cloud computing, the data is stored in cloud server, which needs to be transmitted via network. Under the condition of limited bandwidth, it will bring great delay and quality of service and cannot achieve real-time data processing.
The emergence of edge computing (Satyanarayanan 2017) brings the possibility for researchers to realize massive data real-time system. Edge computing can transfer the computing requirements and storage resources to the edge of the network. By deploying edge nodes, a large number of operations that originally need to be processed by the cloud server can be transferred to the edge nodes, which may greatly reduce the network overhead and cloud performance overhead.
This paper adopts a hierarchical storage model (Zhao 2008) to solve the bottleneck caused by massive data in cloud computing. The hierarchical storage model adopts edge computing to assist data analysis for sensor network. According to the data structure, the hierarchical storage model is divided as edge node layer, cloud service layer, and user application layer. In this paper, the hierarchical storage model is used in a distributed camera sensor network (Adeli 2013) to implement intelligent surveillance system. In the intelligent surveillance system, the edge nodes in edge layer receives the original video from the camera and preprocess the data, such as extracting features of target object. The preprocessed data is uploaded to the cloud server in cloud service layer. The cloud service layer only focus on the integration and analysis of the processed features which are more concise than raw data.