基于模型-模型距离的复杂场景建模及其应用
发布时间：2018-11-26 16:42
【摘要】：背景建模作为目标检测、场景理解等的关键环节,具有重要的研究意义。本文的主要研究内容是基于模型-模型距离的复杂场景建模,主要研究工作如下:针对传统的背景建模算法没有充分利用像素点的时空信息,本文提出了一种基于深度神经网络的背景建模方法。为了更充分的利用像素点的空间信息,同时改变提取空间信息的空间范围,本文设计利用了包含不同膨胀系数的atrous卷积多分支结构,用以提取像素不同邻域范围的空间信息。另外,为了更充分的利用像素点的时间域信息,本文提出了对连续视频帧图像序列进行变间隔采样得到网络输入的方法。通过在数据库ChangeDetection Benchmark Dataset的四个场景中的主观和客观实验对比,本文提出的算法在不同场景中具有较好的适应性,特别是在Office场景中的F-Measure达到92.36%,较经典背景建模算法有提升。针对在前景目标有长时间静止时,前景和背景不能很好的被区分,本文提出了基于模型-模型距离的背景建模方法。本文提出了一种基于深度神经网络的特征提取,并建立特征模型和初始化背景模型。根据深度神经网络能够通过学习不断优化特征的提取方式,使得特征能够更好的表达关键信息,本文通过深度神经网络提取特征。进一步地,本文提出的特征模型中包含了深层网络中的特征和浅层网络中的特征,使得特征模型从不同层面对像素点进行描述。同时利用特征模型对像素点的背景模型进行初始化。为了衡量像素点和其背景模型的相似度,提出了特征模型到背景模型(Feature Model to Background Model,M2M)的距离。针对前景目标突然离开造成的“空洞”,本文利用最小、最大M2M距离对像素点和其邻域像素点的背景模型进行更新。另外,本文介绍了模型更新控制器的自适应更新过程。通过在数据库ChangeDetection Benchmark Dataset的四个场景中的主观和客观实验对比,本文提出的M2M算法较好于主流的五种背景建模算法。特别是在场景Office中的F-Measure达到95.94%,较基于深度神经网络的背景建模算法有提升。
[Abstract]:Background modeling, as a key link in target detection and scene understanding, is of great significance. The main content of this paper is modeling of complex scene based on model-model distance. The main research work is as follows: the traditional background modeling algorithm does not make full use of space-time information of pixel points. In this paper, a background modeling method based on deep neural network is proposed. In order to make full use of the spatial information of pixels and change the spatial range of extracting spatial information, this paper designs and uses the multi-branch structure of atrous convolution with different expansion coefficients to extract spatial information from different neighborhood areas of pixels. In addition, in order to make full use of the time domain information of pixels, this paper proposes a method to obtain the network input by sampling the sequence of continuous video frames at variable intervals. Through the comparison of subjective and objective experiments in four scenarios of database ChangeDetection Benchmark Dataset, the algorithm proposed in this paper has better adaptability in different scenarios, especially in the Office scene, the F-Measure reaches 92.36. Compared with the classical background modeling algorithm, it is better than the classical background modeling algorithm. The background modeling method based on model-model distance is proposed to solve the problem that the foreground and background can not be distinguished well when the foreground target is stationary for a long time. In this paper, a feature extraction method based on depth neural network is proposed, and the feature model and initialization background model are established. According to the depth neural network can learn to optimize the feature extraction way, so that the feature can better express the key information, this paper extracts the feature through the depth neural network. Furthermore, the feature model proposed in this paper includes the features in the deep network and the features in the shallow network, which makes the feature model describe pixels from different levels. At the same time, the feature model is used to initialize the background model of pixels. In order to measure the similarity between pixel and its background model, the distance between feature model and background model (Feature Model to Background Model,M2M) is proposed. Aiming at the "void" caused by the sudden departure of the foreground target, the background model of the pixel and its neighboring pixel is updated by using the minimum and maximum M2m distance. In addition, this paper introduces the adaptive updating process of model update controller. By comparing the subjective and objective experiments in the four scenarios of database ChangeDetection Benchmark Dataset, the M2m algorithm proposed in this paper is better than the five mainstream background modeling algorithms. Especially, the F-Measure in scene Office reaches 95.94, which is better than the background modeling algorithm based on depth neural network.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183
本文编号：2359084
[Abstract]:Background modeling, as a key link in target detection and scene understanding, is of great significance. The main content of this paper is modeling of complex scene based on model-model distance. The main research work is as follows: the traditional background modeling algorithm does not make full use of space-time information of pixel points. In this paper, a background modeling method based on deep neural network is proposed. In order to make full use of the spatial information of pixels and change the spatial range of extracting spatial information, this paper designs and uses the multi-branch structure of atrous convolution with different expansion coefficients to extract spatial information from different neighborhood areas of pixels. In addition, in order to make full use of the time domain information of pixels, this paper proposes a method to obtain the network input by sampling the sequence of continuous video frames at variable intervals. Through the comparison of subjective and objective experiments in four scenarios of database ChangeDetection Benchmark Dataset, the algorithm proposed in this paper has better adaptability in different scenarios, especially in the Office scene, the F-Measure reaches 92.36. Compared with the classical background modeling algorithm, it is better than the classical background modeling algorithm. The background modeling method based on model-model distance is proposed to solve the problem that the foreground and background can not be distinguished well when the foreground target is stationary for a long time. In this paper, a feature extraction method based on depth neural network is proposed, and the feature model and initialization background model are established. According to the depth neural network can learn to optimize the feature extraction way, so that the feature can better express the key information, this paper extracts the feature through the depth neural network. Furthermore, the feature model proposed in this paper includes the features in the deep network and the features in the shallow network, which makes the feature model describe pixels from different levels. At the same time, the feature model is used to initialize the background model of pixels. In order to measure the similarity between pixel and its background model, the distance between feature model and background model (Feature Model to Background Model,M2M) is proposed. Aiming at the "void" caused by the sudden departure of the foreground target, the background model of the pixel and its neighboring pixel is updated by using the minimum and maximum M2m distance. In addition, this paper introduces the adaptive updating process of model update controller. By comparing the subjective and objective experiments in the four scenarios of database ChangeDetection Benchmark Dataset, the M2m algorithm proposed in this paper is better than the five mainstream background modeling algorithms. Especially, the F-Measure in scene Office reaches 95.94, which is better than the background modeling algorithm based on depth neural network.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183
【参考文献】
相关期刊论文 前1条
1 Bi Song;Han Cunwu;Sun Dehui;;Neural network based method for background modeling and detecting moving objects[J];The Journal of China Universities of Posts and Telecommunications;2015年03期
相关硕士学位论文 前2条
1 苏建安;面向智能视频监控的高动态场景建模和修复[D];电子科技大学;2014年
2 陈景东;智能视频监控中的目标检测技术研究[D];华中科技大学;2011年
，本文编号：2359084
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2359084.html
