| As people’s living standards improved, people’s safety awareness is rising. Video surveillance as an effective security tool, are increasingly being installed into various places. And increasing surveillance cameras and more and more clear picture makes streaming media server overwhelmed. In order to solve the contradiction between the rapid growth in the number of high-definition cameras and the single server limited service ability. This paper discusses parallel algorithm of the target detection and tracking based on distributed cluster.First, study and analyze the existing detection and tracking algorithms, verify the detection efficiency of Gaussian mixture model, and found it difficult to ensure real-time processing of HD video.Secondly, improve and decouple algorithm of using the Gaussian mixture model to detect target, put forward sub-block detection thinking, which split each frame image into multiple parts to process by different units. For different processing unit capacity in the cluster, I raised a thought that several pixels merge into small pieces and use the average of pieces to make model. For the reason that stagnant target will be trained to the background, I put forward the cumulative amount of difference to determine whether the target stagnation.Furthermore, in order to increase the role of the newly measured data filtering, improved the Kalman filter by adding a gradual elimination factor. Use Kalman prediction to preliminarily classify targets to different processing units histogram matching. For multi-camera matching problem, a path model is raised to improve the matching efficiency.Finally, the improved and decoupled algorithm was written into Topology program which is run in Storm clusters, to verify the detection and tracking results and analyze the real-time perform. We found that the parallel algorithm of target detection and track based on Storm clusters can fully guarantee the real-time requirements. |
PDF Full Text Request