Research On Moving Objects Detection With Complicated Condition

Moving object detection is a challenging and very important topic in computer vision field, which covers computer science, optical, mathematical, control science, cognitive science and so on, and has been widely applied in our real life, such as military, industry, medicine, meteorology, traffic management, public security, etc. so research on motion detection has important theoretical significance and practical value. However, so far there are a number of urgent practical problems need to be solved, especially in complex background environment or dynamic scenarios. Generally, video surveillance environment, including indoor and outdoor variety, a scene often composed by background, moving object (such as pedestrians, vehicles), disturbance objects (such as the shaking of the branches, swinging curtain) etc, which has many actual problems, such as the background changes caused by illumination changing, misjudgment caused by disruptors disturbance, motion state changing, high noise in the background region, difficult extracting of moving targets'effective features, lack of imaging devices, real-time monitoring, etc, therefore, there are still challenges and difficulties in practical applications. As a result, how to improve the intelligent degree of video processing, expand its range of application and improve the performance of system has become a hot spot in research and application field.In this work, we mainly research and discuss several key technologies of moving object detection, such as optical distortion, environmental illumination changing, shadows and camera movement, which often occurs at image sequences captured by monitoring equipment.For real-time optical distortion correction issue of video capture camera, GPU-based real-time nonlinear geometric correction algorithms for the large field image is proposed, which not only meets the monitoring system on-line processing demands, but also greatly eliminates the negative impacts of the wide-angle lens barrel distortion on the target Extraction; Through the theoretical analysis of camera calibration model and non-linear distortion model, the transformation matrix and solving methods of distortion correction are abstracted and concluded; The problem of reduce the amount of data at the expense of accuracy is solved by combining the GPU hardware and CUDA software architecture, which make full use of its parallel computing advantage; while taking into account the various factors leading to lens distortion, so a number of lens distortion coefficient, which contain the radial distortion, centrifugal distortion and thin prism distortion coefficient, are solved by using the least square method, and the use of cubic interpolation algorithm to non-integer coordinates of pixel interpolation, greatly improving the calibration accuracy.For classification feature selection issue, the effect of pixel characteristics on detection algorithm performance in different color space is systematic analyzed, which provides a theoretical basis and practical guidance for color feature selection and algorithm optimization under typical environments. For color images, in addition to brightness information can be used in gray-scale image, but also chroma, saturation, etc, are increased, which can improve the moving object detection capability by focusing on the development and take full advantage of color information. At present, the most moving object detection algorithm choose the color characteristics as testing standards, but different color space reflects different image information. Therefore, we analyze the color space models and the information changing of moving objects in different color space, for example, RGB, YCbCr, HSV and Lab color space; Background subtraction results, which base on Gaussian mixture model algorithm with different colors spatial characteristics under complex background environment, are compared; Experimental results shows:in RGB, YCbCr, HSV and Lab color space, YCbCr, and Lab are better in many respects, such as anti-interference of background light, environmental adaptability, robustness.For illumination changing and shadows interference issue, color feature is combined with local binary pattern features of the unified model as a texture feature, and Choquet fuzzy integral is introduced, a new moving object detection algorithm is proposed based on multi-feature similarity fusion operator, background/ foreground classification model by membership degree is established, a new idea about high-precision object extraction in complex environment is opened up. The choice of classification features directly affects moving object detection results. Color is one of the most commonly used features, but the instability of natural background determines that color is sometimes difficult to accurately describe the image information, and usually the methods with individual classification are sensitive to the dynamic changes of the scene, such as illumination changing, shadows and reflected light, etc., which makes the moving object extraction inaccurate. Therefore, we combine color features with block-based texture features to improve the foreground/ background classification accuracy, through the link-by-analysis of feature selection, similarity measure definition, multi-feature fusion as well as the segmentation threshold determining, a moving object detection algorithm based on the color-texture fuzzy integral feature is proposed. In the feature selection aspect, the combination mode of color features and texture features is selected. To reduces the illumination changing on the impact of moving object detection, YCbCr color space, which separates brightness from color, is adopted to describe the color characteristics; The rotation invariance of the LBP with the extended form of a unified model of operator ULBP K.R ri is defined to describe the texture feature, which not only has advantage of non-parametric express and non-monotonic changes in gray-scale effects, but also makes histogram entries compact, and not susceptible to noise, thus avoiding interference of illumination changing and shadows; In multi-feature fuse aspect, Choquet integration method from fuzzy mathematics is introduced, which integrates similarity measure of the color and texture features, to improve the classification accuracy rate; In segmentation threshold determination aspect, an adaptive strategy is defined, in the foreground/background segmentation stage, classification threshold adaptive solving step is increased, which makes the threshold T c.t (x, y) no longer invariable, but change as environmental change, and the frequency and step length are set as adjustable parameters to improve the convergence speed and solution accuracy.For dynamic Scene issue, image registration method of dynamic background based on the arithmetic mean of multi-scale Harris corner point square-root is proposed, the measurement model of camera movement is established, the motion compensation is completed, finally moving object detection of dynamic scene is achieved by three frame differencing. In practical application, the moving camera led to aliasing between the changes of background and changes of moving objects in the image. To eliminate the disturbance, a new moving object detection algorithm of dynamic scenes based on Harris corner points'SAM is presented. First, register the current frame image with the previous image to catch the global motion parameters; compensate image motion by using of the motion parameters; and then take frame differencing method with the first two images, to get the outline of moving object; regard moving region as mask, and detect and locate the moving target. In image registration aspect, an image registration algorithm based on the SAM of multi-scale Harris corner points is proposed. Local invariant point, which usually use for image retrieval and identification, is introduced into the image registration field. To improve the repetition rate of the matching points, multi-scale Harris corner detection operator, which easy-to-extract and anti-disturb of images translation, rotation, brightness change, etc, is used; To obtain stable points, the angle, scale and location of the candidate matching points are iterative screened by clustering and SAM information; During the parameters solution, only stable corners are matched to reduce the computation, and no need to optimal search, which avoid falling into local extreme; In moving object segmentation aspect, three frame differencing method is used, namely, movement area is obtained by calculating the difference values of the adjacent three consecutive frames, the motion range of moving targets between adjacent frame is quickly detected by thresholding the images'differencing; To improve the accuracy of motion detection, adaptive thresholding is used for difference image binary. The experimental results show that this algorithm can deal with positioning of moving targets of dynamic scenes, such as camera motion, etc. which lay the foundation for moving target tracking and identification;For Infrared images, in accordance with the imaging characteristics of infrared image, the moving object detection algorithm of the visible light image sequences is extended to this area, infrared object detection algorithm based on combination of luminance information and texture information is proposed, the similarity fusion based on fuzzy integral is extended to the infrared image motion detection. Infrared imaging system becomes hot research in the intelligent visual surveillance field for its advantages of all-weather working, no shadows, anti ambient light disturbance, etc, but there are also unique challenges, such as low signal to noise ratio, polarity reversal, the halo effect and so on. For infrared imaging features, image de-noising and enhancing algorithm based on correlation of wavelet coefficients is used to enhance noise suppression to solve the contradiction between noise suppression and image enhancement; moving object detection algorithm based on fuzzy integral of the color-texture features is extended, A new background modeling and moving object detection Algorithm for infrared image is proposed, namely, infrared image moving object detection algorithm based on the fuzzy integral of brightness-texture features. The Sugeno fuzzy integral theory is introduced into the infrared target detection, the combination of brightness features and texture features is regarded as classification characteristics, fuzzy measure and fuzzy integral are used for classification, which effectively resolve the limitation exist in traditional background differencing method used in infrared image.The moving object detection under complex background is a very important and challenging research field, this research work plays a role in promoting this field research, the put forward and improved method offers a number of new way for moving object detection technology, experiments test and verify that this paper's work will enrich and push forward the studies of the related areas in both theoretical and technological aspects.