Research On Optimal Algorithms For Video Moving Object Detection

Video moving object detection is widely used in video surveillance. And the performance ofthe intelligent video surveillance system will be directly affected by the detection performance. Inthis thesis, we mainly focus on improving the detection performance from two aspects, research onthe optimization algorithm based on the background subtraction method and research on the Graphcut optimization algorithm based on the interested region extraction. And the contents summarizedas following:Among the methods for the moving objects detection, background subtraction is preferred.And this method mainly contains two steps background modeling and foreground detection.Gaussian mixture model is widely used in the background modeling based on the backgroundsubtraction detection，while the value of the learning rate do limit the robustness and the update rateof the model during the detection process. So in this section, we aim to improve the detectionperformance by optimizing the background model parameters. During the specific optimize processwe construct a new background model based on the goodness-of-fit, combined with the Gaussianmixture distribution and the Empirical distribution. And then we combined with the mathematicalderivation and the nonnegative matrix iteration method to optimize the model parameters to get theoptimal model and complete the subsequent model updates. The updating process of the modelparameters proposed in this paper only depends on the statistical features of the samples to bedetected. So the optimized algorithm well avoids the problems caused by the learning rate in theGaussian mixture modeling, and improve the performance of the detection indirectly.While for a moving object detection the Graph cut algorithm is another method that has beenused. And the high computational complexity, low detection efficiency and poor adaptability are allbig challenges for the Graph cut algorithm. To improve the detection performance based on theGraph cut algorithm, we proposed the research on the optimal Graph cut algorithm. This algorithmproposed a cascade cutting theory, which do the preliminary cutting based on the likelihood ratiocriterion, to complete the extracting of the region of interest (ROI) to reduce the region to bedetected，and in the ROI we do a second cutting based on the graph cuts algorithm to extract themoving objects, the preliminary cutting well reduce the amount of calculation and improve thedetection efficiency. Meanwhile to improve the adaptability of the detection，we combined theKalman filter with the geometric characteristic extraction to realize the estimation of the modelparameters, which well estimate the following flow parameters to be detected, and is helpful to thesubsequent detection based on the cascade cutting, and well improve the performance of the detection.In the experimental section, we choose groups of video clips under four different conditionsincluding the detection under simple or complex background, indoor or outdoor, with illuminationchange or no illumination change, a single or multiple targets. Finally, we combined with theamount of the experimental simulations and the corresponding experimental data analysis to verifythe advantage and applicability of the algorithm optimize in this paper, which do improve theperformance of the video moving object detection.