Background Modeling And Target Detection In Complex Dynamic Scene

Moving object detection is an important aspect of computer vision. Backgroundsubtraction is one of the commonly used methods of moving object detection. The key ofbackground subtraction is background modeling. A good and effective backgroundimage is the safeguard of accurately detecting moving object. Background modelingmethod has the assumption of a static scene usually, but the actual background ofabsolute rest is difficult to obtain, the weather, periodic motion of the object, camerashake and other reasons would break the assumption of a static scene. Therefore, thestudy of background modeling method in dynamic scene has important significance.This thesis in-depth studies the background model of Gaussian and nuclear densityin the dynamic scene. In order to get better object detection results, this thesisintroduces shadow elimination method based on the background model to correct theerroneously detected object and improve the object detection performance. The maincontents of this thesis as follows:(1) Research three simple and common background models: the frame difference,timeline weighted filtering and Kalman filtering. The principle and specific processes ofmethod are described in detail in this paper. Through simulation experiments, theperformance of the three models were compared and analyzed, and in accordance withtheir limitations, this thesis leads to the text of the Gaussian distribution model andkernel density estimation model.(2) The thesis discusses in detail the background model of mixture Gaussiandistribution and kernel density estimation, the former belongs to the parameterestimation method, the latter belongs to the non-parametric estimation methods. For themixture Gaussian model, the research focuses on the problem of parameter update. Forthe nucleation density, it focuses on the formula meaning and the effect of windowwidth.(3) The thesis introduces a shadow elimination method based on the overall edgeinformation, and details the process of this method. This method is combined with the background subtraction to correct the rough foreground objects to get more effectivemoving target.(4) For several methods mentioned in this paper, experiments are implemented.Through the experiments, we analyze and discuss the advantages and disadvantages ofeach method performance.