Adaptive Moving Target Extraction Algorthm Based On Visual Background Extraction

Moving target detection technology of video is an important branch of current computer vision.It is widely used in many fields such as intelligent surveillance video and military equipment.At present,there are many mature target detection algorithms.Background subtraction is the most classic method,which has attracted the attention of scholars.The background subtraction method is divided into direct subtraction method and background modeling method.Compared with the background subtraction method,the background modeling method is better.Among the common background modeling methods,the performance of the visual background modeling method is the best,but the detection effect is not accurate enough in complex situations.This thesis is mainly based on the visual background modeling method,and conducts in-depth research on video moving object detection in complex environments.The following is the specific research contents.(1)Firstly,this thesis studied the video background extraction algorithm,made theoretical introduction and experimental analysis of the commonly used background extraction algorithms,and summarized the characteristics and applicable scenarios of various algorithms,then based on the shortcomings of these algorithms,this thesis proposed a new background extraction algorithm.Then it researched the video foreground target detection algorithm,made theoretical explanation and experimental analysis of the current foreground target detection algorithms,and summarized the characteristics and applicable scenarios of various algorithms.Finally,the background modeling method in the background subtraction method is introduced.The commonly used background modeling methods are introduced and analyzed in detail.Based on the best performance based on visual background modeling method,its deficiencies are addressed.In this thesis,an improved foreground object detection algorithm based on visual background modeling is proposed.(2)Secondly,in view of the shortcomings of the existing background extraction algorithms,there are low timeliness when there are many video frames,target smears in the extraction results when there are foreground targets in the video,and dynamic backgrounds as interference factors in the scene.There are many problems such as noise in the extraction results.This thesis proposed a background extraction algorithm based on pixel interval weighting.The algorithm first divides the gray value into ten intervals,and then counts the first three intervals where each pixel in the image appears most frequently in all video frames,and then takes the first frame of the video as the research object.If the value belongs to one of the three intervals,the pixel value in the first frame is taken as the pixel value of the position in the background image,otherwise the weighted sum of its corresponding three intervals is taken as the pixel value of the position in the background image.Because the algorithm not only considers the foreground target in the video,but also takes into account the dynamic background interference factors in the video,and statistics between pixel partitions,greatly reducing the time complexity of the algorithm on the premise of improving performance.Experimental results show that the algorithm performs better than background extraction algorithms.(3)Finally,this thesis analyzed the foreground object detection algorithm based on visual background modeling.For the presence of foreground objects in the first frame of the video,smear phenomenon will appear in the detection results,and when there is a dynamic background in the video scene as an interference factor,the problem of a large amount of noise in the extraction results was studied.Through analysis,it is found that the smear phenomenon is mainly caused by the inaccurate initialization of the background model of the visual background modeling.Therefore,the background image extracted by the background extraction algorithm based on the weighted sum of the prime points is used as the model based on the visual background modeling method.The initial value solves the smear phenomenon caused by the presence of a foreground target in the first frame of the video.The existence of noise in the detection results is due to the use of a global fixed threshold for the pixel segmentation strategy.Therefore,this thesis proposes a matching algorithm based on adaptive thresholds.The algorithm determines the dynamic degree of the current pixel based on the distance between any pixel in the background model of each pixel,and then calculates the current matching threshold based on the dynamic degree.When the degree of dynamics is large,the dynamic background can be detected as the background pixel ranks by appropriately increasing the matching threshold,otherwise the matching threshold can be reduced so that the front spots with small changes can be detected.This thesis selected videos in different scenarios in the CDnet 2014 public dataset to conduct simulation experiments.Through quantitative analysis and comparison of the quantitative analysis index accuracy and recall rate of foreground target detection,the algorithm of this thesis detects performance in various complex scenarios,both are better than other conventional foreground target detection algorithms and have better robustness.