Research On The Key Techniques Of Large-scale Video Surveillance Analysis Based On Streaming Computing

Large-scale intelligent video analysis has become the urgent need of video monitoring, but the two problems are faced in the specific implementation: one is the accuracy of the intelligence algorithms. An accurate and efficient algorithm can reduce false positives and error rate, and make an effect truly intelligent monitoring. The other is the construction of analysis platform. How to solve the computing power required, IO and other issues during a large-scale video analysis. This article settled in analysis platform-based streaming and discussed how to solve the computing power and IO faced other issues of large-scale and real-time analysis of surveillance video by means of stream computing.We design and implement a large-scale surveillance video analysis system based on stream computing. Three aspects of the design and implementation of the system mainly are focused on: 1) To avoid the bottleneck of the IO. Compared with the small file of log files, the amount of surveillance video streaming data is bigger, so the first answer how to transfer large-scale monitoring flow into the analysis platform. 2) How to organize large-scale surveillance video analysis needed computing power. Intelligent video analysis algorithms are computationally intensive work and need to consume large amounts of computing resources. 3) To achieve computing scalability to meet user’s changing requirements for video analysis scale. Our solution is based on the SPARK analysis platform, which is the integration technology of stream computing and in-memory computing, and optimized the SPARK to make it suitable for large-scale intelligent video analysis of this application scenario. Finally a prototype system is developed to verify our ideas.In summary, the main work of this paper is as follows:First, in the large-scale video streaming together when imported into the analysis platform, we adopt a method based on the computing power aware for receiver location allocation. This strategy can be based on the ability of compute nodes to disperse the video stream IO on different hosts, solving IO distribution problem.Second, in the realization of real-time processing of video streams of data, this paper use the real-time task scheduling strategy based on local-data. This paper employs the improvement of real-time task scheduler algorithm based on local-data for fast processing of data. The data access based on memory computing technology can effectively reduce the IO access time and improve the processing efficiency of large-scale video analysis, realizing effective transformation the real-time value of video streaming.Third, in terms of computing scalability, this paper discusses the adaptive computing frame dropping strategy to ensure the stable operation of the system. After horizontal scalability of the system is discussed, this paper adopted a method for dynamically adding computing nodes to increase computing capacity of the system.Preliminary research indicates that the approach for large-scale analysis of surveillance video streaming based on stream computing is feasible. This approach can make large-scale video network monitoring system work and achieve the effect of real-time intelligent analysis of large-scale video.