Research On Adaptive Mixed Gaussian Target Detection Method

In recent years,computer vision technology has been widely used in various fields,such as intelligent monitoring system,traffic detection,automatic guidance and so on.As the moving target detection technology is the key and basic work of computer vision technology,that is,the movement target detection will directly affect the next key links,such as identification,classification,trac king,and behavior analysis,so the study of moving target detection is very necessary.Most of the current target detection methods are based on background modeling methods.the background modeling method based on the hybrid Gaussian model is used commonly.The method can handle multi-model scenes,such as foliage swaying and fluctuating surfaces.However,in the actual target detection pr ocess,due to the impact of light changes,the method can not update the background model adaptively,At the same time,the moving target will cause shadow,resulting in the test results are not accurate.In addition,since the method itself has a fixed nu mber of models per pixel,causing a large amount of computation.Simultaneous detection of the moving target will produce smear phenomenon.Aiming at the above problems,the main contents of this thesis are listed as follows:1)The change of illumination is divided into global illumination and local illumination.This thesis presents a measure of the change in the degree of global illumination and the introduction of light parameters to determine the degree of local light changes,Through the combination of the two to distinguish the different changes in light,and then use the adaptive learning rate to meet the various changes in light to achieve the adaptive update of background model.The experimental results show that the method can accurately detect the moving target under the illumination change.Moving objects under light will produce shadows,and simple and efficient methods are based on chromaticity methods to remove shadows.But when the background with the background color similar to the shadow ca n not be removed easily.Because the gradient feature has the invariance of illumination,a chromaticity method based on the fusion gradient feature is proposed,and the moving target is extracted by combining the chrominance information with the gradient information.Compared with the chromaticity-based method,the experimental results show that the improved chromaticity-based method can better remove the shadow and improve the accuracy of the detection.2)Pixels usually use a fixed number of Gaussian mod els to describe its state,resulting in a large computational problem.Because the number of Gaussian models is related to the threshold of background selection,the threshold is self-adaptive,and a threshold method based on Tsallis entropy is proposed.The problem of following contour caused by moving objects is due to the fact that the traditional hybrid Gaussian model algorithm only learns in the time domain without taking into account the relationship between the pixel and the surrounding pixels.Theref ore,we propose a parameter to determine the smear,that is,the relevant proximity.For the fuzzy background points to introduce the variable v for adaptive adjustment.The experimental results show that the smear can be removed very well.