Research On The Detection Technology Of Moving Object In Video Sequences

As one of the most important technologies in computer vision, the moving object detection andtracking technology which based on video sequences is widely used in all kinds of automationsystems. This thesis mainly researches on the moving object detection technology based on videosequences.This thesis analyzes three classical methods used in moving object detection technology basedon video sequences, including optical flow method, frame difference method, as well as thebackground difference method, and the strengthes and weaknesses,as well as the applications ofthese methods are disscussed. On the basis of VC++6.0simulation platform including the OpenCVopen-source library, the dectection result of background difference method based on statisticalaveraging and Gaussian mixture model is studied respectively in the indoor and outdoorsurveillance video. In indoor video whose main noise is caused by light changes, the backgrounddifference method based on the statistical averaging model has a better detection performance thanthe method based on Gaussian mixture model. However, in the surveillance video of outdoorparking lot whose noise may caused by more complex conditions, the result of the backgrounddifference method based on Gaussian mixture model is more accurate than that of the method basedon statistical averaging model.Compared with the background difference method based on Guassian model, backgrounddifference method based on codebook model is better able to deal with the complex environmentbackground, effectively restrains noise disturbance and has good real-time performance. But thetraditional codebook model after training is not changed, unable to effectively respond toenvironmental changes and the target contour extracted by this method is not accurate enough.According to the problem mentioned above, an improved method based on codebook model todetect moving object is proposed in this thesis. By using a cache codebook, realize the real-timeupdate of the codebook when detecting the moving object and furthermore use the updatedcodebook to do the detection of moving object. Finally, the fusion of results detected by three framedifference method and background codebook model difference method is done. The methodproposed in this thesis can effectively reduce the error detection when the environment changes andincrease the accuracy of extracting the target contour by doing the experimental simulation.