Foreground Detection And Application In Multiple Environments

The detection and tracking of foreground targets under surveillance video has always been the focus of digital image processing,machine vision and other fields,and it is also difficult.Special foreground target detection is more important,and it is the basis for tracking and identifying future prospective targets.The effect of foreground detection directly affects the effect of later tracking and recognition.Therefore,this paper mainly studies the foreground target detection in multiple environments,and studies the traditional algorithms from the principle,and analyzes their advantages and disadvantages through experiments to improve them.Firstly,this paper summarizes the foreground detection process and model under the current surveillance video,so as to conduct special research on different models.Among them,the foreground detection algorithm is mainly divided into frame difference method,background modeling method,optical flow method and deep learning method.This paper mainly studies the first two types of related algorithms.For the frame difference method,this paper compares the adjacent interframe difference method and the three frame difference method,and finds that the three frame difference method has better detection effect than the former;for the background modeling method,this paper studies the traditional background difference method and Gaussian mixture.The model,and also the moving object detection algorithm based on inter-frame difference and background difference and the VIBE algorithm are studied.It is found through experiments that the algorithm effect of the background modeling class is generally much better than the frame difference method in the detection of dynamic and static background.However,the frame difference method is much faster than the background modeling algorithm in terms of detection speed,and has better real-time performance.However,these two types of algorithms show very poor prospects for abrupt transition from static state to motion state.The detection effect,especially the background modeling algorithm,will have the problem of "ghosting" detection.Secondly,this paper proposes a foreground detection algorithm based on improved three-frame difference method and improved Gaussian mixture model for the problem of "ghosting" detection.The three-frame difference method is added to the eight-field information to improve it.The Gaussian mixture model is added with statistical methods to improve its initialization,and then the EM algorithm is used to quickly iterate to optimize the parameters.The foreground region detected by the mixed Gaussian model is matched with the foreground region maskdetected by the improved three-frame difference method,and the matching foreground is reserved.For the foreground region that is not matched,the foreground region is regarded as a 'ghost' region,and the pixel corresponding to the current frame is used.Dot Gaussian distribution model update,replace the mean ?of the maximum weight Gaussian model,and quickly reconstruct the background,thus solving the traditional method of detecting holes and detecting 'ghosting' problems.The experimental results show that the proposed algorithm has better robustness and accuracy in different backgrounds,and the accuracy and accuracy are better than traditional algorithms.Finally,this paper uses the strong real-time performance of the improved three-frame difference method to determine the foreground target region.Through the connectivity analysis,template A is used to initialize the manual cutting target target of the Mean-Shift algorithm,and the matching condition and the correction condition are added.It is found through experiments that the improved Mean-Shift algorithm can achieve automatic single target and multi-target tracking to some extent.