Research On Video Computing Architecture And Method For Large Scale Surveillance System

To realize the concepts behind the term "safe city" and "smart city",video surveillance technology has received a great attention by the research community.In order to improve the efficiency of the surveillance system and realize intelligent surveillance,the video computing technology for the large-scale surveillance system was proven to be critical.This paper focuses on studying the bottleneck of data transmission,computing ability and the high false alarm rate of video analysis algorithms,hence,these issues were being faced when a large-scale video analysis system is in consideration.Considering the bottleneck of data transmission,A marginalized video computing architecture is proposed,which deploys and allocates computing resources around the video data source,in order to overcome the problems faced when large amounts of video data are being transmitted.By deploying regional computing nodes to process video data and core node to unify the scheduling and manage regional computing nodes,the video computing resource deployment and the distribution mode based on the region of the video source are being realized.The computing nodes send the calculated information fragment to the core node,which can effectively reduce the communication bandwidth consumption.Considering the bottleneck of computing ability,a dynamic adjustment strategy of computing resources is proposed.Considering the low value of the video,an adaptive video frame dropping strategy and a computing resource allocation method based on the value of video are proposed.When the computing task is increasing and the computing power is insufficient,the video frame of some cameras will be discarded according to the frame discard strategy,which will reduce the cost of computing resources,to ensure the stability of the system and the computing tasks with high urgency will obtain sufficient resources instantly.Considering the problem of high false alarm rate in video computing,combined with the pedestrian detection algorithm,which is commonly used in video computing,two methods are proposed,in order to reduce the false alarm rate and they are as follows: 1)Adaptive Pedestrian Detection Method Based on Credibility.The credibility of the pedestrian detection classifier is divided into two dimensions(i.e.time and space),and the feedback of the judgment result is introduced to update the reliability.So that the false alarm rate gradually reduced.2)Correction Method of Pedestrian Detection Results Based on The Bayesian Algorithm.The Bayesian algorithm is used to train the historical judgment results,and a modified classifier will be formed.The modified classifier will be used to make the second judgment on the recognition result of the detection classifier.Based on the aforementioned calculation architecture and methods,a video computing platform for large-scale surveillance system is designed and implemented.The proposed system was applied in many "safe city" systems,which insures its feasibility,effectiveness and practicability.