基于随机森林算法的目标检测与动作识别方法研究

发布时间：2018-04-02 11:04

本文选题：目标检测　切入点：Haar方向梯度描述子　出处：《吉林大学》2016年硕士论文

【摘要】：随着世界上科技创新能力的不断提高,带有非现代化标签的传统视频监控技术已经不能满足人们的需求,如何更智能化地完善监控系统成为重要的研究课题。这其中涉及了很多图像处理的研究技术,比如对视频画面中运动目标进行实时捕获、建立分析模型使监控系统理解画面中场景内容,以及更高级地对人类异常行为发出自动警告等,这些研究内容促使更多学者投身到计算机视觉领域,为工业信息现代化作出了极大的贡献。着眼于现有的海量视频数据信息在图像领域有巨大的研究价值,本文利用随机森林这一机器学习技术手段,提出基于运动目标检测与人物动作类型识别的新型算法,与众多仅停留在图像像素处理层面的内容不同,本文赋予模型对目标的分类判别能力,利用深度挖掘的图像特征信息、有监督学习架构及多种图像技术手段来完成。在目标检测方面,特征信息对于目标的表述尤为重要,本文在前人的基础上对方向梯度直方图这一特征向量进行了补充,引入Haar小波工具,将图像的高低频信号提取出来,再经由方向梯度直方图将其串联,形成Haar方向梯度描述子;接着,对霍夫森林算法的理论、应用以及不足之处做出陈述,重新构造随机树节点的分裂准则,在文本分类信息上验证模型的有效性;然后,依据图像块提供的特征信息,重新定义随机树增量式的生成方法,获取节点的投票图像,应用Meanshift算法取得概率密度极值,对含有目标信息内容的区域进行有效分割提取;最后,通过引入两个规则对分类器的分类效果及图像的检测准确度进行评定,经由7组实验分析验证算法的准确性。在动作识别方面,本文通过三个方面进行概述,分别为特征信息表现层、模型构建层及动作分类层。在特征信息表现层中,针对原始光流场难以表述人物运动信息的问题,本文在原有亮度恒定假设基础上,又增加了灰度强度梯度不变、速度分量场区域平滑、区域描述子匹配连续性和相邻帧描述子一致性的假设条件,再对计算后的光流场以色彩信息编码,将目标的运动信息表现出来,实验结果表明这种光流算法能够描述运动物体的变化信息,为后续两层提供必要内容;在模型构建层,本文从人物动作在时空域下具有连贯信息的角度出发,提出一种以图像立体块为基础的模型构建方法,在原有的2维图像上增加时间信息,结合节点的判别标准,以增量生长的形成方式重新设计随机树;在动作分类层,依据叶节点投票生成的动作空间图判断人物动作类别。本文对3组数据库分别进行了实验,通过混淆矩阵及比较动作识别准确度来验证算法合理性。
[Abstract]:With the continuous improvement of the ability of scientific and technological innovation in the world, the traditional video surveillance technology with non-modern labels has not been able to meet the needs of people. How to improve the monitoring system intelligently has become an important research topic.This involves many research techniques of image processing, such as real-time capture of moving objects in video images, establishment of analysis models to enable monitoring systems to understand the scene content in the pictures,As well as more advanced automatic warning of human abnormal behavior, these research contents urge more scholars to devote themselves to the field of computer vision and make great contributions to the modernization of industrial information.Aiming at the great research value of the existing massive video data information in the field of image, this paper proposes a new algorithm based on moving target detection and character action type recognition, which is based on the machine learning technology of stochastic forest.In the aspect of target detection, feature information is very important to the representation of target. In this paper, the feature vector of gradient histogram is supplemented by introducing Haar wavelet tool to extract the high and low frequency signal of image.Then the directional gradient descriptor (Haar) is formed by concatenating it with the directional gradient histogram. Then, the theory, application and deficiency of the Hoff forest algorithm are described, and the split criterion of the random tree nodes is constructed.The validity of the model is verified on the text classification information, and then, according to the feature information provided by the image block, the incremental generating method of random tree is redefined, and the voting image of nodes is obtained, and the probability density extremum is obtained by using Meanshift algorithm.Finally, two rules are introduced to evaluate the classification effect of the classifier and the accuracy of image detection, and the accuracy of the algorithm is verified by seven groups of experiments.In the aspect of action recognition, this paper summarizes it in three aspects, namely, the representation layer of feature information, the layer of model building and the layer of action classification.In the representation layer of feature information, aiming at the problem that the original optical flow field is difficult to express the movement information of people, based on the original assumption of constant brightness, the grayscale intensity gradient is invariable, and the velocity component field is smooth.The region descriptor matches the hypothesis of continuity and the consistency of adjacent frame descriptors, and then encodes the calculated optical flow field with color information to show the moving information of the target.The experimental results show that this optical flow algorithm can describe the changing information of moving objects and provide necessary content for the following two layers.This paper presents a model construction method based on image solid block, which adds time information to the original 2D image, and redesigns the random tree in the motion classification layer by incremental growth, combining with the discriminant criterion of nodes.According to the action space graph generated by the leaf node voting, the character action category is judged.In this paper, three groups of databases are experimented to verify the rationality of the algorithm by confusing matrix and comparing the accuracy of action recognition.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】