摘要
运动目标检测在智能视频监控、人机交互、目标导航等诸多领域有着广泛应用.背景减法是运动目标检测中应用较广泛的一种方法.在该方法中,背景建模和阈值化分割是最重要的步骤,直接决定了检测效果的好坏.当目标本身变化比较大时,若利用传统的基于全局阈值的分割法,分割效果并不理想.针对基于全局阈值分割差分图像存在的问题,本文提出了一种基于混合高斯背景模型的像素分类运动目标检测方法.该方法首先利用混合高斯模型对背景建模,克服了场景变化等因素带来的影响;其次,利用背景减法得到差分图像并对像素进行分类,最后对分类后的像素集分别进行阈值化分割,得到前景目标.实验结果表明,与传统的基于全局阈值的分割法相比,本文算法能够获得更好的检测效果和鲁棒性.
Detection of moving objects(DMB) is widely used in many areas,such as intelligent video surveillance,human-machine interaction and target navigation.The background subtraction algorithm is widely used for DMB.In the traditional algorithm,background image is generated from the reference frame and is abstracted from the current frame to form differential image.Afterwards,the moving regions are separated from the differential image by predefined threshold.The procedure of generating background image is called background modeling and the procedure of separating moving objects is called moving regions segmentation.These two procedures are so crucial that they are related to the detection performance directly.However,when detecting the objects with large variation,most of the time,this traditional algorithm based on global threshold suffers from performance degradation.The reason is that the correctness of both background modeling and global threshold are no longer assured.In order to solve this problem,an object detecting algorithm using pixel classification based on Gaussian mixture model is proposed.The Gaussian mixture model,whose parameters and weighting factors are updated on-line,can overcome the influence of scenes change,such as branches swaying and illumination changing.In the proposed algorithm,the usage of Gaussian mixture model to complete background modeling is the first step.Secondly,the pixels of the differential image,which is obtained by subtracting the extracted background from the current frame,are separated into two parts by the predefined threshold,whose value is set by experiment.Because separating into two parts appears to have satisfied experimental result,considering the balance of both complexity and performance,the number of separated parts is set as two.Owing to different characteristic of these two parts,their front images are generated separately.The method to produce the front image is to compare the pixel value with different thresholds in different parts.The algorithm is implemented in Matlab 7.0,and two standard test sequences from the CVRR(Computer Vision and Rototics Research) database in UCSD(University of California San Diego) are used to verify the effeteness of the proposed algorithm.Comparison is made between the traditional algorithm based on global threshold and the proposed one based on pixel classification.Experimental results show that the proposed algorithm can achieve much better detection rate and lower false alarm rate than the traditional one.In terms of the comparison of subjective result,the proposed algorithm obtains the outline of the moving objects clearly,which proved that the robustness of the proposed one can be guaranteed.
出处
《南京大学学报（自然科学版）》
CAS
CSCD
北大核心
2011年第2期195-200,共6页
Journal of Nanjing University（Natural Science）
基金
广西自然科学基金(桂科自0991058)
广西高校人才小高地建设创新团队资助计划项目(桂教人[2007]71号)
关键词
背景差分
混合高斯模型
像素分类
运动目标检测
background subtraction,Gaussian mixture model,pixel classification,moving objects detection
