Research On Video Object Structure And Re-recognition In Multi Camera Environment

With the rapid development of smart city construction,more and more network video surveillance devices are added to every corner of the city,and massive surveillance video data are generated.These video data can not rely solely on the human eye to complete the monitoring,so it is urgent to develop a more intelligent network video monitoring system.The video data structuring is one of the key technologies to solve the intelligent analysis of video data.Video data structuring includes video moving target detection,video target association,video data structured description,storage and management and other key technologies.The re recognition of video target of cross camera is one of the key technologies of intelligent monitoring of video surveillance network.In this paper,video target detection,video target association,video data structured description,storage and management,and target recognition in Cross video camera are studied.First of all,we analyze the existing video structuring programs based on Gaussian Mixture target detection.For the limitations of the scheme,this paper propose a video structuring scheme based on deep learning target detection.Secondly,Since the low accuracy and high failure rate problems in deep learning target detection model of specific targets detection,we design an enhanced data set about pedestrian and vehicle target in surveillance video,and train YOLOv2 neural network model using the specific data set,the detection result of pedestrians and vehicles target shows higher performance.Third,based on the data of structured video object detection of deep learning method,this paper proposes a correlation algorithm of discrete targets based on probability correlation matrix,the algorithm combined centroid distance feature,color histogram and SURF features each other and convert as three probability matrixes,and make a fusion by weighting the importance of the feature.Experiment shows an excellent video discrete target association.Fourth,based on the data of video targets association sequence after the deep learning targets detection process,this paper proposed a describe scheme of video sequence structured data using the scalable XML and a structured data storage scheme,and design a structured target data unified management system.Finally,based on the video target description,storage and unified management,this paper puts forward two kinds of pedestrian re-identification algorithms using video target sequence to process enhance pedestrian surveillance video data intelligently.one of algorithms calls pedestrian sequence re-identification based on small histogram variance,and the other calls pedestrian sequence re-identification based on region segmentation feature.The former uses the overall color feature of the video target sequence,while the latter takes advantage of the local significant color features of the video target sequence.The algorithm in this paper has strong correlation.The final video target recognition experiment confirms the effectiveness of the preceding detection algorithm and association algorithm one by one.