Research On Moving Target Detection Algorithm Under Dynamic Background

Foreground detection in video surveillance is a core issue of computer vision technology for decades.The accuracy of subtraction can significantly affect the performance of the following high level computer vision tasks,such as moving object tracking,human activity recognition and scene analysis.Consequently,moving objects detection in video surveillance has been a hotspot to domestic and overseas scholars.Recent research on problem formulations based on decomposition into foreground plus background matrices shows a suitable framework to detect moving objects from the background.The basic operation of foreground detection in video surveillance consists of separating the moving objects called "foreground" from the static information called "background".Many effective and efficient researches in Background/foreground separation have been proposed based on this methodology.However,according to the characteristics of foreground and background in the real-world video surveillance,background model of the scene usually constructed with disturbance factors,such as dynamic illumination,moving objects,fuzzy picture and so on,moving objects in the foreground can be size-changed and texture-changed.Objects detection is still a challenging tack when it comes to dynamic backgrounds and size-changed foreground objects.Based on the frameworks of the foreground detection and background subtraction,three moving objects detection methods in video surveillance are proposed:1.An adaptive local low rank model for object detection.Signals exhibit different structure at different scales,we proposes a low rank model for background subtraction which defines an adaptive low rank decomposition system with adaptive partition model and adaptive regularization parameter setting.It performs the matrix decomposition of video datasets by block operation which is based on adaptive saliency-based size to separate the object to its corresponding scale.The method is mainly comprised of adaptive modeling part and low rank approximation part.The adaptive modeling part constructs adaptive block matrices with saliency-based matrix partition.The blocks in different size from input image matrix will be extracted and reshaped to obtain the block matrices to be of low rank approximation.Rank reduced SVD of the block matrices which are partitioned from the input image with adaptive size helps to extract interested objects.By this mean,we can obtain target efficiently.2.A multi-scale low rank model with Wasserstein distance for object detection.Wasserstein distance,which called Earth-Mover distance as well,is introduced to construct cost function for which can effectively capture both the similarity and morphologic structure of the probability distributions.In particular,it can help the algorithm be stable to tolerate sudden background variations like the changing background without losing sensitivity to detect moving objects.An adaptive generalization of multi-scale partition model was proposed by representing the data matrix as a sum of block-wise low rank matrices with adaptive saliency-based scales of block size.It can obtain an accurate and compact representation of scenes in video surveillance with multi-scale structure.A Least Information Entropy(LIE)criterion is given to select the target among the multi-scale low-rank approximation,which would have the least clutter and the most information of foreground moving objects.This model provides remarkable improvements to adapt the complex background and foreground moving objects in real-world datasets with infrared small target.Experimental results demonstrate that the presented method performs well in video surveillances with dynamic background.3.An adaptive Gaussian mixture model of background separation for object detection.The classical Gaussian mixture model preserves signature of the background information by constructing reliable model based on mixed Gaussian.Since it required a flexible and self-adapted system for the dynamic background,we construct an real-time updated mixed Gaussian model based on the cost function of Wasserstein distance,which calculate the similarity judgment between different probability distributions can be applied to the meaningful corrections in background model.This model provides remarkable improvements to adapt the complex background and foreground moving objects with changing size in different scenes.