基于概率假设密度滤波的图像多目标跟踪与轨迹保持方法研究

发布时间：2018-04-01 12:06

本文选题：视频图像　切入点：目标检测　出处：《哈尔滨工程大学》2016年硕士论文

【摘要】：视频图像中目标检测与跟踪属于机器视觉中的前端处理部分,通过对视频图像信息进行处理,实现计算机能够代替人眼进行目标的识别与跟踪功能,一直是机器视觉和计算机视觉的最终目的。如何从图像序列中获取有用的信息作为科研人员的研究方向已成为视频图像中目标检测与跟踪的理论基础。视频图像目标检测与跟踪已大量应用于现实生活中的各种场景,在这些方面都显示出对视频图像信息处理的重要应用价值。本课题主要关注视频图像中多目标跟踪与轨迹保持方法,对图像序列进行多目标跟踪与轨迹保持等方面研究。第一,在由检测到跟踪方法中,图像目标检测是进行目标跟踪的前提。使用基于图像的颜色、梯度方向和梯度幅值等基本图像特征的积分通道融合特征,结合基于boosting的决策树分类器,生成目标检测器,实现对图像目标的准确检测。第二,以自顶向下的处理思路,阐述了基于贝叶斯滤波框架的目标跟踪原理。研究基于随机有限集理论的概率假设密度滤波递归算法,对算法的两种不同实现方式进行对比仿真实验,最终确定使用高斯混合PHD滤波算法来进行图像多目标跟踪。第三,结合图像目标表观模型特点和高斯混合PHD算法的应用条件,提出适合用于图像多目标跟踪的高斯混合PHD自适应新生目标算法,通过在标准行人检测数据集上实验,验证跟踪算法的性能。第四,针对PHD滤波处理并不能实现多目标跟踪运动轨迹情况,构建图像多目标能量函数,将视频多目标跟踪问题转化为能量函数优化求解问题。把经过跟踪滤波算法处理的图像序列作为输入,进行多次迭代,实现视频图像多目标跟踪轨迹。
[Abstract]:Object detection and tracking in video images is a part of front-end processing in machine vision. By processing the video image information, the computer can replace the human eyes to recognize and track the target. It has always been the ultimate goal of machine vision and computer vision. How to obtain useful information from image sequences as the research direction of researchers has become the theoretical basis of target detection and tracking in video images. Target detection and tracking have been widely used in various scenarios in real life. In these aspects, it shows the important application value to video image information processing. This subject mainly pays attention to the method of multi-target tracking and track keeping in video image, and carries on the multi-target tracking and trajectory keeping to the image sequence and so on. First, Image target detection is the premise of target tracking in the detection and tracking method. The integral channel fusion feature based on the basic image features such as color, gradient direction and gradient amplitude is used, and the decision tree classifier based on boosting is combined. Generation of target detector to achieve accurate detection of image targets. Second, to top-down processing ideas, This paper introduces the principle of target tracking based on Bayesian filtering framework, studies the recursive algorithm of probability hypothetical density filtering based on stochastic finite set theory, and makes a comparative simulation experiment on two different realization methods of the algorithm. Finally, Gao Si hybrid PHD filter algorithm is used for image multi-target tracking. Thirdly, considering the characteristics of the image target model and the application conditions of Gao Si hybrid PHD algorithm, Gao Si hybrid PHD adaptive newborn target algorithm suitable for image multi-target tracking is proposed. The performance of the tracking algorithm is verified by experiments on standard pedestrian detection data set. Fourth, In view of the fact that the PHD filter can not realize the multi-target tracking, the multi-target energy function is constructed. The video multi-target tracking problem is transformed into the energy function optimization problem, and the image sequence processed by the tracking filtering algorithm is taken as the input, and the multi-target tracking trajectory of the video image is realized through several iterations.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】