Semantic Video Moving Objects Detection And Tracking Based On Cauchy Distribution

This thesis focuses on the rigorous and robust algorithms for both detecting and tracking semantic video moving objects within cluttered scenes using a monocular video sequences, which high level analyses of intelligent visual surveillance can be found on. Meantime, the algorithms can also be applied in a wide variety of fields, such as tele-medicine, content-based coding and description, content-based video indexing and browsing, interactive video, entertainment and so on.The intelligent visual surveillance has emerged as an important and active topic in image processing and computer vision, which is also one of the most challenging problems due to many difficulties still unsolved. Because of the variety of scenarios, a surveillance system has to behave differently according to the particular application. The monitored scenes, even in controlled environments, are undergoing continual change. Changes including the environment's illumination, shadows, reflections, sensor artifacts and etc. can change the overall appearance of seemingly "static" scenes from fixed cameras. Natural motions, such as moving clouds and tree branches pose additional difficulties for objects detection and tracking. Presently, there have been many visual surveillance systems developed with a variety of software and hardware architecture. However, moving objects detection and tracking are not only two key technologies for all of the surveillance systems, but two primary handicaps to improve the performance of a surveillance system. In the case of the current stat of art, it is still urgent affairs to explore the robust and efficiently computed algorithms for both rigid and no-rigid moving objects detection and tracking in complex changing scenes.Based on the principle of intelligent visual surveillance and the connotation of semantic moving objects detection and tracking, a detailed review of their current state of art is discussed. In this thesis, an anatomy of change detecting method used to detect moving objects is proposed according to its processing scheme, e.g. feature extraction, feature analysis, pixel classifying and post processing.The methods of background modelling and maintaining are presented. According to the signal characteristic of video image and the need to cope with both color and gray images, YCbCr color space is selected as the description of the pixel of a background model, which separates the brightness from the chromaticity component.The background model can be initialized by using previous empty scene background images, or by an automatic restoring program. In order to be adapted to varieties of background changes in a real applying environment, the thesis puts forward a method to update the background model using three different levels, e.g. pixel level, region level and frame level. This method utilizes the feedback from moving objects detection and tracking so as to make the process of background update be a closed loop with negative feedback.For the first time, this thesis put forward a novel method of background modelling and subtraction based on a local-linear-dependence-based Cauchy statistical distribution model. A ratio of pixels intensity or color between two images or two difference images is used as the feature for subtracted background modelling. We find and demonstrate that the statistic distribution of the ratios of the pixel's intensity value between two frames of background images obeys Cauchy distribution, as well as the ratio of a pixel's intensity value is local-linear-dependence to all or part of its nearby pixels assume that there hasn't been a change or they are belong to a same object in the two images. A robust background modelling and subtraction approach based on this is acquired. The Cauchy based method without exponential operation is more cost-efficient than the Gaussian's. At last, Experimental results demonstrate the proposed algorithm can tolerate the whole or local sudden or slow change in illumination, filter clutter noise caused by small motion in scene background, and ad...