|
Visual analysis of objects motion is one of the most important researchtopics in the domain of computer vision;it is also an active field which hasinterested many researchers in recent years. Objects motion analysis aims atdetecting, identifying and tracking objects from image sequences, further,describing and understanding their behaviors. Motion detection belongs tothe low-level vision problems, is the basis of other object motion analysis.The technique of motion detection has already obtained a substantialprogress through several years’ research, but the complexity and multiplicityof the supervising background makes missed detection and false alarms stillpresent, so motion detection can’t satisfy people completely and so calls forfurther research.Some common approaches of motion detection are introduced in thispaper, including the approach based on character, the approach based onoptic flow, the approach based on frame subtraction and the approach basedon background modeling. Their advantage and disadvantage also with thefield they fit are pointed out too. We should adopt the proper motiondetection algorithms for specified applications. In this paper, the applicationis an indoor surveillance. In this case, the video is captured with a fixedcamera, and the background’s change is slow. The algorithm present in thispaper has high real-time quality and availability, can detect small motionobjects, and is fit for real-time detecting and tracking system.The correlation coefficient testing algorithm present in this paper isbased on a fact that in the region of foreground, the vector of pixel intensityacross multiple frames tends to be highly correlated with its spatialneighbors, and vice versa. Noise vectors are less correlated with their nearneighbors than signal vectors. Therefore we can decide if a pixel belongs tothe foreground by the intensity of correlation, at the same time we canremove noise effectively.First, we define pixel vector :which is composed of the intensity valuesof pixels at the same coordinates in all given group of frames, then followthe define in Probability, formally define the concept of correlationcoefficient, and according to this formal definition, we design a new methodfor motion detection. This approach compares the changing pattern of a pixelin the time domain with its neighboring pixels, and computes the intensity ofcorrelation by the formal define of correlation coefficient, further decides ifthe pixel belongs to the foreground. The steps of correlation coefficienttesting algorithm are as follows:1、First, construct pixel vector for the pixel at location(i,j):Where fi,(jn+k-1) denotes the intensity value of(i,j) in frame k, and Ndenotes the number of given frames.2、For every pixel in a given neighborhood centered at (i,j) and not locatedat(i,j),compute the pixel correlation coefficient as defined. Then computethe maximum value of the pixel vector correlation coefficient(denoted asM(i,j))within the neighborhood of pixel (i,j).3、Compare M(i,j)with a given threshold thr, if M(i,j)>thr, then label(i,j) as apixel of foreground, otherwise, background. The binary value is saved asthe result of motion detection, Change Detection Mask (CDM). Thr ispredetermined, in our implementation it is the optimum value gained inseveral experiments.4、Perform postprocess including removing noise and filling, then separatethe moving object from background according to CDM.This paper also considers the infection of shadows caused byillumination. Shadow is the translucent district in the picture, the cover ofobjects makes the intensity value in the background changed. Such regionwill be detected as foreground when without processing. In this paper,combined with shading model, we carry out a series of operations on theintensity values before correlation coefficient testing method, remove factorof illumination, and thus avoid the infection of shadows.The result of motion detection may be not the real moving region, theremay be some region caused by noise, and for some objects, maybe just theiroutline is detected, the internal region which has even quality can’t bedetected, so the postprocess is needed. In consideration of demands of highrobust of small change and real-time quality, the postprocess in this paperadopt an easier strategy differing from the traditional method usingMathematical Morphology operations: all holes should be filled;any islandwhose size is larger than a predetermined quarter is retained as foreground,otherwise removed as noise.In postprocess, filter in broad sense is carried out firstly, this isaccording to the erode operation in Mathematical Morphology, but theselection of the structure element is easier. Then a connected componentsalgorithm is applied to mark each separated”1” region. Lastly fillingoperation is performed to gain the final CDM. Through these steps smallnoises are removed and the holes in the moving objects are filled too.According to the forenamed correlation coefficient testing algorithm,we construct a motion detection model, and act it on a series of testingsequences and some video segments which imitate the input of camera totest this algorithm .The results show that this model can accomplish thepre-establish target of the experiment successfully: Detect moving objects’presence or leave availably, in the meantime fully considered the validity incase of complicated background and not rigid body;compare theexperimental results to other methods’ at the same time, the contrast resultsshow that, the method present in this paper is highly robust in detectingsmall moving object than other methods, and also can effectively andexactly detect moving foreground objects, this method has lowercomputational complexity, so is fit for real-time transport.
|