Moving Objects Detection In Video Based On Graph Cuts

Object detection technology is one of the hot problems in the research of the computer vision.It concerns many fields such as video image process, artificial intelligence and pattern recognition,and has wide application in video supervision, traffic management, medical treatment, imagecoding, human-machine interface and so on.As the disadvantages of background subtraction is easy affected by the sunlight change, andoptimizing the problem of over-smoothing of the video moving objects detection, we presentthree optimization approaches to detect the video moving objects based on the introduction ofgraph cuts. Graph-cut algorithm is a energy minimization method based on Markov random fieldand maximum a posteriori. In this paper, we improve the graph-cut algorithm by the characteristicof moving objects and the higher-order Markov random field.In this paper, an adaptive graph-cut algorithm based on ROI (region-of-interest) wasproposed. The nodes flow of graph-cuts algorithm are adaptive updated based on the Kalmanprediction of the number of moving objects pixels and objectives-background pixel-pairs. Theextraction of ROI including the moving objective not only can effectively reduce the computationof graph-cuts algorithm, but also can improve the detection accuracy of the moving objects.The traditional graph-cut algorithm to detect video moving objects is based on the low-orderMarkov random field. In this paper, two algorithms to detect video moving objects based onhigher-order Markov random field are proposed. First, we optimize the node flux by theintroduction of elasticity energy in order to get an accurate segmentation. The Euler’s elasticamodel is one such higher order model of central importance, which minimizes the curvature of alllevel lines in the image. The Euler’s elastica model can preserve discontinuities. As a result, thedetection target will have a more natural appearance. Second, one optimization detection methodbased on higher-order Markov random field is proposed. This algorithm can detect movingobjects adaptively for estimating the area, perimeter and shapes of moving objects. Processing isbased on Kalman filter theory and dynamic model of motion image situation. The higher-ordermodel is used to restrict the shape of moving object.The graph-cut algorithm based on ROI is adaptively adjusted by the kalman prediction toimprove the accuracy of segmentation. The extraction of ROI is used to eliminate backgroundinterference and to reduce the computational time. The graph-cut algorithm based on higher-orderMarkov random field can improve the accuracy of details which consider the relationships ofenergy among multi-pixels. Not only these optimization methods can improve the detection accuracy in static background conditions, but also the quantitative detection indicators canperform very well in complex background conditions such as shadows cast by moving object,illumination variations, dynamic backgrounds and so on.