Image Feature Extraction And Classification And Recognition Based On Video Object Retrieval

Computer technology is the typical representative of today's rapid development ofscienceandtechnology, especiallywiththecomputervisionbecomingmoreandmoremature, the computer demonstrates its unparalleled ability, making it a better way forpeopletoobserveandunderstandthings. Intelligentvideosurveillancetechnologyisahot research of computer vision, which is equivalent for computers installing "eyes"to capture the required video, and then the captured images are transferred to thecomputer.Finally,thepowerful computingabilityofcomputers is usedtoanalyzeandunderstand the objects in the images. The studyof intelligent surveillance has a lot ofpractical significance. It is free from the limitations of the traditional manualmonitoring, and it can monitor the object for the analysis of understanding, not onlysaves a lot of manpower and material resources, but also brings greater economicprofits, which plays an important role for social and economic development, as wellasscienceandtechnology.The working process of intelligent monitoring system is: firstly, we use the videocapture devices to capture images in the form of frames. Then, codec technologytransfers the video stream to the computer terminal. Thirdly, we use motion detectionto extract the foreground objects and extract features and track the target in thedynamic situation. Finally, we achieve the functions of object classification andbehavior understanding. This paper mainly studies image's feature extraction,classification and recognition based on the retrieval of video stream. Under thepremise of well extraction of moving object, the study focuses on the foregroundfeature extraction and description, and uses support vector machine classifieralgorithm to classify objects roughly and determines the optimal combination ofcharacteristic values. Finally, we use three kinds of constraints and specific situationtofurtheridentifymovingobjects.First of all, Gaussian mixture background model and background subtraction areused to extract the moving target. Based on the characteristics of the experimentaldata, this paper uses morphological analysis principle to pre-process the objects aftersegmentation,but in practice, due to the interference of the external environment, as well as video capture quality issues images are often addressed a lot of noise, whichcan affect not only background but also foreground and reduce the completeness andaccuracy. This paper firstly filters parts of noise in the background by use of contourarea threshold, and then by adjusting the template size and shape, as well as openingoperation and closing operation, we finish filling empty objects, smoothing borders,removing the burr, etc. Experiments show that there is little deformation offoreground, and which is very close to real situation. But when there is a lot ofshadow and the color of background and foreground is very close, even afterde-noising,theprospectsofanobjectwillstillbeseverelydeformed.Secondly, in the process of describing foreground objects'feature, since the colorandtexturecharacteristics oftheobject don't containtheinformation oftype,theyarenot appropriate for object classification and recognition. This paper adopts the shapefeatures both in time and in space to describe the foreground objects. Space feature isfor single static image, reflecting the inherent characteristics, such as size, locationand other information characteristics, and time feature is for continuous motionimages, reflecting motion changes in spatial features, such as speed and direction. Onthe existing basis, there are four characteristics: width height ratio, space room ratio,velocityandrateofchange inanobject the size. This paperproposes two newfeatureextraction methods: one is the ratio of the local width of moving target and overallwidth; the other is the ratio of minimum distance and maximum distance from thecenter to the contour. Experiments show that the former feature can not onlyeffectively separate people and cars, but also have good stability. At the same time itis also an important method for vehicle identification. And this paper improves thespace room ratio method, we use the ratio of the target area and its outer leastrectangular area. Experiments show that taking multi-angle of objects into account,thisvalueofratioismoreclosetothetruevalue.Thirdly, in the research of object classification, this paper uses support vectormachine as a classifier, and chooses the radial basis function as the kernel function.We use LibSVM software training samples under cross-validation methods to selectparameters of penalty factor and the kernel function, and then we use this two bestparameters selected to train samples and get a model of SVM. Finally, we use thismodel to test the samples. This paper uses five different combinations for the sixproposed features .Experiments prove that the use of support vector machines forclassificationofmovingobjectsinthevideocanbeahigherrecognitionrate. Finally, on the basis of simple moving objects classification, this paper makes afurther study of recognition of behavior, and presents a characteristics descriptionapproach based on three levels: the description of objects feature information, timeconstraints and spatial location. Thus, different constraints are proposed tocorresponding behavior based on the above three aspects, including the fast-movingpedestrians, object detained, object moved, cars stopping in certain areas and peopleassembling. We achieves the goal byfinishing a simple video-based system, which isabletodosimplevideosurveillancetasks,includingmovingtargetdetection,tracking,classificationandrecognitionofbehavior.To sum up this work, this paper mainly uses the shape-based approach to extractmoving target features, and uses support vector machine to achieve the classificationforobjects,aswellastheoptimizationandselectionofthefeaturecombination.Thereare many limitations in the realization and there is still a lot of further work to do todeviseasystemthatissmarterandcanbeapplicabletoallsituations.