Research And Implementation Of A Parallel Retrieval System In City Video Surveillance

With the construction of “Safe City” project, the city video surveillanceinfrastructure has been widely deployed in large and medium-sized cities, which playan important role in the bus conductor, criminal investigation, public safety, cityinformation statistics, and many other aspects. Due to massive and unstructuredfeatures of these real-time video surveillance data, the work of video data analysis,such as the target object retrieval, need to rely on a lot of manpower. To solve thisproblem, a video retrieval system based on MPI parallel framework, object-orientedobjectives is designed and implemented. The main contents are as follows:A moving target detection module is designed. It can realize the functions likemoving target detection and segmentation of images from the video database. Thisarticle analyzes the advantages and disadvantages of existing moving target detectionalgorithm, and designed some parallel optimization strategies to effectively improvethe speed of codebook background model training. Meanwhile, a moving foregrounddetection algorithm based on hybrid strategy is proposed, which combines theadvantages of Gaussian mixture algorithm and codebook algorithm to improve thedetection accuracy of the moving target. Under the experimental environment whichconsists of four MPI parallel computing units, the codebook background modeltraining parallel speedup is greater than3.9. Detection algorithm based on hybridstrategy achieved higher detection accuracy than single algorithm. Finally,experiments proved the availability and efficiency of moving target detection module.A color-based target retrieval module, which can extract color features of eachimage of the moving object in the dataset, and conduct similarity measure with theinput image, is designed. This article analyzes the research status of existingcolor–base indexing methods, and designs to use color correlograms for imageindexing and comparison. Further more, some parallel optimization strategies forextracting color correlograms are proposed. Under the given experimentalenvironment, the parallel speedup of our color-correlograms extraction algorithm isup to3.0. Experiments proved that the target retrieval module can obtain accuratesearch results. Based on MongoDB database, a set of data structures, which can couple the twocomputing modules efficiently, is designed. The strategies of constructingsemi-structured target database and its implementation is studied. Meanwhile, thearticle describes an index construction algorithm based on locality-sensitive hashing,which projects high dimensional features of similar images onto the same bin inlinear Hamming space to improve the retrieval performance for massivehigh-dimensional data.The parallel retrieval system proposed in this article has good structure and clearprinciples. According to input image, it can complete those boring jobs like findingsimilar moving objects in massive videos intelligently. Without loss of accuracy ofthe algorithm, the parallel architecture can significantly improve system performance.