The Algorithms Of Intelligent Video’s Multiple Target Detection Based On The Gaussian Mixture Model

Intelligent video technology includes moving target detection, recognition and classification, tracking, behavior understanding and description. Among them, moving target detection is the foundation and premise. Of course, there are a number of background modeling method by using different background model and updating methods, such as Median filtering, W4,Single Gaussian Model, Background statistical model, Codebook, Senior background statistical model, Nonparametric background modeling. Background subtraction is the common method used to detect moving target, that is, to conduct the threshold error between the background without target and the current image. The paper present an improved multiple-target method which uses spatial information to compensate time information. The random number generation method is used for sampling the neighborhood to complete background modeling the spatial distribution of the pixel during the time distribution based on Gaussian mixture background modeling. Meanwhile, the foreground detection algorithm which is based on pixel’s history statistic information and decision fusion mechanism to get a more accurate extraction of the judgment in static and foreground moving target. The shadow detection scheme which based on the exploitation of the HSV space and the connected components is also introduced in this paper. Besides, the principle work finished as follow:(1) The research on moving target detection algorithms. The paper make a comparison between several ordinary method used to detect moving target, study their fundamental theory and make a description about their advantages and deficiencies. To be frank, the comparison concentrate on the amount of calculation, the result of detection, the resistance to background disturbance and storage space requirements. Meanwhile, in order to solve the problem that dynamic background and the change of light will disturb the multiple-target detection, a modified algorithm for Gaussian mixture model is proposed in paper. The algorithm will be used to background model and improve the robustness of dynamic background owe to the strategy that spatial information and time information compensate each other.(2) The modified algorithm for Gaussian mixture background. After implementing the moving target detection based on Gaussian mixture model, we present an improved algorithm to eliminate the deficiencies of typical Gaussian model, for example, the bluntness to sudden movement and mutations of light. The soul of the method is to make pixels’life cycle adapt to actual distribution by the pixels’time and space distribution characteristics so that keep the consistency of space and rebuild the covered background to meet the requirement of precision. On the other hand, Gaussian mixture model only make use of information on time distribution. Therefore, it is dull to sudden changes in the scene. So it’s susceptible to dynamic background and updates slowly. But these problems can be solved by background model on space distribution.(3) Image post-processing. To eliminate the serious impulse noise and salt and pepper noise by median filtering algorithm can get a better result than mean filtering and the others. The paper also achieve a satisfied result by shadow suppression algorithm based on HSV color space transformation. In addition, in order to improve the result, it fill up the part hollowed out in the detected moving target by operation of mathematical morphology.The theory and result make it clear that the algorithm is robust to dynamic background and have a high accuracy. What’s more, it can meet the demand of real-time. When it comes to the moving target static and shadow detection, the demand can be satisfied by this method. And relying on connected domain processing, it can get the whole mask image and improve the veracity and integrity of the outline. In the experiment of foreground detection in the dynamic background, the method(the improved GMM) detect the Standard Data Set people(FPR is0.015), the Recall is89.1%, better than the others. And its recall rate beyond the mean filtering54%. The running time is0.0303s/frame. And its Recall is93.8%when it detect the Standard Data Set office(FPR is0.015), still higher than the others. It is74.7%higher than Frame Difference and mean filtering.The running time is0.0469s/frame.