Research And Implementation Of Moving Object Detection And Tracking Algorithms

Vision is an important means for humans to observe and understand the world. Alongwith the development of information technology, it has become an aspiration of humans toendow machines with the function similar to human vision, so computer vision technologyemerged as the times required and has developed rapidly. Intelligent visualsurveillance(IVS) is an emerging applied research orientation in the field of computervision.Basedonthetraditionalvideosurveillance, IVSrealizestheprocessing,analysisandunderstandingof videosignals usingcomputer visiontechnology, andthen theanalysis andunderstandingresults can be used to control the video surveillance system itself, so that theintelligentized level of the video surveillance system is raised. The core technologies ofIVS mainly include moving object detection, object tracking, object classification andrecognition, object behavior understanding and some other aspects. IVS technology can beapplied into public safety monitor, traffic management, military reconnaissance, industrialprocesscontrol,medicalcare,celestialobservation,customerserviceandmanyotherfields.It has great practical value and broad application prospect in both civil area and militaryarea. In IVS systems, moving object detection and tracking are the most basic intelligentanalysistechnologiesthatareappliedmostwidely,andthereforeareofgreatsignificancetotheintelligenceofsurveillancesystems.The main tasks of the movingobject detection and tracking based on video stream areas follows: processing the video image sequence, and segmenting the foreground objectregions accurately, completely and clearly in every frame of video image; computing theposition of each object in every frame of video image, that is to say, locating each movingobject in the video stream continuously to offer the motion trajectory of each object, andcomputingthevelocity, accelerationandothermotionparameters of eachobject integratingthecorrespondingkinematictheories.This paper researched and improved moving object detection and tracking algorithmsbased on video stream through studying some existing moving object detection andtracking algorithms, image processing, Gaussian model and Kalman filter theories, andreferencing a lot of literatures inside and outside country. Experiments showed that good effectshadbeenachieved.First of all, several commonly used moving object detection algorithms wereresearched, analyzed and improved. Existing moving object detection algorithms aremainly classified into three categories: optical flow, frame difference, and backgroundsubtraction. These algorithms have their own advantages and disadvantages. For themoving object detection in video surveillance systems with fixed camera, backgroundsubtractionisthemostcommonlyusedmethod.Theideaofbackgroundsubtractionmethodis to compare the current video image frame with a background model, and the regions ofpixels that have larger difference are considered to be moving objects. This paper deeplyresearched several commonlyusedbackgroundsubtractionalgorithms: medianfilter,singleGaussian model and Gaussian mixture model, analyzed their modeling principle and set uptheir algorithm flow charts. Detailed experimental analysis and theory discussion wereperformedtodeterminethelengthoftheslidingstoragewindowformedianfilteralgorithm.Strategies were put forward to balance the contradiction between the detection effect andthe time and space complexity through discussing how to choose an appropriate length ofthe sliding storage window for different video segments collected under differentcircumstances. An improved background modeling algorithm based on Gaussian mixturemodelwasproposed.Gaussianmodelisdeterminedbytwoparameters:themeanvalueandthe variance. Different learning mechanisms for the two parameters have direct influenceon the stability, accuracy and convergence of the model. According to the differentcharacteristics of mean value and variance, we adopted different learning rates for them.For the mean value update equation, we adopted the learning rate put forward by theliterature[36].Forthevarianceupdateequation,weadopted aconstant as thelearningrate.Thiskindoflearningmechanismcanmakethemeanvalueconvergequicklyandaccurately,andcanalso makethevarianceconvergequicklyandsteadily.Experimental results showedthe better foreground object detection effect of the improved background modelingalgorithmbasedonGaussianmixturemodel.Secondly, comparisons were drawn between several commonly used moving objectdetection algorithms. Their object detection effect, memory requirements, real-timecapability,algorithmcomplexity,noisesensitivity,adaptabilitytocomplex environmentandsome other performances were summarized and compared through a large amount ofexperiments. Experimental results showed that the frame difference method is simple andfast, and occupies very little memory, but has bad detection effect, which is applicable tothe cases that need high real-time capability and low detection accuracy. The median filter method and the single Gaussian model method reach good detection effect under relativelysimple environments, but show unsatisfactory performance under complex environments.They can't adapt to busy surveillance scenes, and can't deal with environmentaldisturbances such as wind blowing the trees, fluttering flags and shaking of camera, whicheasily lead to false detection, so the two algorithms can't satisfy the applicationrequirements of complex environments. Besides, the median filter algorithm requires muchmemory space. In contrast, the Gaussian mixture model method showed much betterperformancethanthepreviousthreemethodsundercomplexenvironments. Itcandealwithrepeatedfluctuations withlimitedrangesuchas windblowingthetrees,flutteringflags andshaking of camera. The Gaussian mixture model method has good detection effect, properalgorithm complexity, relatively low memory requirements and good real-time capability,so it is asignificant methodincomplex backgroundmodelingfield. However,whendrasticchanges occur in the surveillance environment, such as turning on or off the lights, darkclouds covering the sun and so on, except frame difference method, all the other threemethodscan'tadaptsoon.Thirdly, an image post-processingdenoisingmethod based on binarymorphologywasproposed according to the specific application requirements and the features of theexperimental data in this paper. In practical applications, due to the disturbances of theoutside environment, the reasons of the camera itself and so on, the video images collectedalways havemanynoises.Thenoises reducethequalityoftheimages,blurtheimages,andtherefore have a seriously bad influence on the object detection effect, so it is a necessaryand critical step to eliminate noises. Because the object detection results need to be passedon to other project team members as input data for the research on object classification andrecognition, an image denoising method that not only can eliminate noises in backgroundregions but also can retain the integrity and accuracy of foreground regions to the greatestextent is required in order to offer as much efficient information as possible for the objectclassification and recognition. So in this paper, the conditional dilation operation wasadopted first to eliminate noises in background regions efficiently, and then the closingoperationandopeningoperationwereintegratedtoperformoperationssuchasfillingholes,smoothing borders and removing burrs and little bridges in foreground regions.Experimentalresultsshowedthegoodeffectoftheproposedimagedenoisingmethod.At last, a fast and efficient method for moving object detection and tracking was proposed. Real-time capability is very important in IVS systems. Especially in somesituationsthatneedhighsafetyrequirements,timelyandeffectivedetectionplaysakeyrolefor the reliability of the system. In addition, the higher the video image resolution is, themore abundant and the more accurate the image information detected is, and therefore thegreater the help is provided for the following operations such as object recognition andearly-warning decision-making. But higher video image resolution will bring more timespending, which has a bad influence on the real-time capability of the system. Consideringthe importance of real-time capability for IVS systems, a feedback model was proposed inthis paper to decrease the computational cost of the moving object detection and trackingoperations, and hence the real-time capability of the system was improved. This feedbackmodel was proposed on the basis of the moving object detection algorithm based onGaussian mixture model and the moving object tracking algorithm based on Kalman filter.Kalman filtering is recursive, which has two steps: predictingthe current state based on theprevious state, and using the current observations to update the prediction. So theprocessing steps of the feedback model are as follows: First, state prediction directsattention to regions of interest. Next, pixels within the regions are processed to generateobservations. Finally, the observations drive state estimation. Experimental results showedthat theproposedmethodhas goodobject detectionandtrackingeffect,markedlydecreasesthe computational cost and enhances the real-time capability without influencing theaccuracy. So the proposed method is of great significance to the practical applications ofIVSsystems.To sum up, this paper did some shallow researches on moving object detection andtracking algorithms based on video stream, proposed an improved background modelingalgorithm based on Gaussian mixture model, performed in-depth discussion, summary andcomparison of some commonly used moving object detection algorithms, proposed animage post-processing denoising method based on binary morphology, and proposed a fastand efficient method for moving object detection and tracking based on Gaussian mixturemodel and Kalman filter. Along with the development of technology and the progress ofsociety,movingobjectdetectionandtrackingtechnologywillgrowincreasingly.Ihopethatthe work done in this paper can make slender contribution to the development of movingobjectdetectionandtrackingtechnology.