【作者】 刘颖；

【导师】 汤春明；

【作者基本信息】 哈尔滨工程大学 ， 信号与信息处理， 2012， 硕士

【摘要】 细胞运动追踪的研究一直是生物学、生物医学和细胞学的重要组成部分。尽管活跃细胞在序列图像中所占的比例不大，但它却有着非常重要的作用。而准确地预测和追踪序列图像中的活跃细胞仍然是一个尚未解决的难题。建立细胞的运动模型并找到一种准确有效的细胞追踪算法是本文课题研究的重点。本文提出了利用扩展卡尔曼粒子滤波器（EKF-PF）对序列图像中的活跃细胞进行预测和追踪，该算法主要是适用于非线性非高斯的系统，而序列图像中细胞的运动多是非线性非高斯的，尤其是活跃细胞，所以利用该算法进行预测和追踪会更加符合活跃细胞的运动特点。通常在细胞追踪时大多会采用匀速或匀加速运动模型，但活跃细胞的运动并不是简单的匀速或匀加速运动，所以本文提出了机动加速度的非零均值时间相关模型来模拟活跃细胞的运动模型，该模型所描述的运动是介于匀速运动和匀加速运动的一种形式，更加符合活跃细胞的运动特点。由于一般的运动模型只是利用目标的位置、速度和加速度的信息建立的，并没有考虑目标的运动角度信息。为了更加准确有效地预测和追踪活跃细胞，在建立运动模型时加入细胞的运动角度信息来对模型加以改进。同时在建立细胞的代价函数时，通常是利用细胞运动距离的变化、运动速度大小的变化、直径大小的变化、面积大小的变化、方位角大小的变化和离心率大小的变化这六个准则建立代价函数，为了减少运算量，本文只采用细胞运动距离的变化、直径大小的变化和方位角大小的变化这三个准则来改进细胞的代价函数。通过对粒子滤波器的基本算法和几种改进算法的性能比较分析，本文最终选择了EKF-PF算法来预测和追踪活跃细胞，这主要是因为该算法充分利用了新的观测值,使得重要性概率分布更接近后验概率分布，能更加准确有效地预测和追踪序列图像中的活跃细胞，同时算法运算量不大，能较好地满足预测和追踪的实时性。本文对三组序列图像中的14个活跃细胞进行预测和追踪的仿真实验。实验结果表明序列图像中的活跃细胞的估计值和观测值的误差都在2.5个像素以内。因此这种EKF-PF算法基本能够实现对活跃细胞准确有效地预测和追踪。更多还原

【Abstract】 The cells motion tracking study has been an important part of Biologi, Biomed andcytology. Although active cells are a very small proportion of cells in image sequences, it hasa very important role. Active cells’ accurate prediction and tracking in image sequences is stillan unsolved problem. Establishing cell motion model and finding an accurate and effectivecell tracking algorithm are the key of subject study in the paper.The paper applies Extended Kalman Particle Filter (EKF-PF) to active cells predictionand tracking in image sequences. The algorithm mainly applies to the nonlinear andnonGaussian system, and cell motion in image sequences is mostly nonlinear andnonGaussian, especially active cells. So the algorithm which applies to predict and track ismore conformed to active cells motion characteristics. Usually cell tracking mostly adoptsconstant velocity or constant acceleration motion model, but active cells motion is not simpleconstant velocity or constant acceleration motion. So the paper present non-zero mean valuetime-correlation model of the maneuvering acceleration to simulate active cells motion model.The motion which the model describes is between constant velocity motion and constantacceleration motion, which is more conformed to active cells motion characteristics. Thegeneral motion model only uses the target position, velocity and acceleration information toestablish, which did not consider the target motion angle information. In order to predict andtrack active cells more accurately and effectively, cell motion angle information is added toimprove the motion model when it is established. We usually use six criteria, which are cellmotion distance change, cell motion velocity change, cell diameter change, cell area change,cell azimuth change and cell centrifugal rate change, to establish the cells’ cost function. Inorder to reduce the calculated amount, we only use three criteria, which are cells motiondistance change, cells diameter change and cells azimuth change, to improve the cells’ costfunction.Through the performance of the particle filter fundamental algorithms and several kindsof the modified algorithms for comparative analysis, we finally choose the EKF-PF algorithmto predict and track active cells, mainly because the algorithm takes full advantage of the newobservations, the important probability distribution is more close to the posterior probabilitydistribution, which can more accurately and effectively predict and track active cells in image sequences. At the same time the algorithm calculated amount is not too large, which can betterto meet the real time of prediction and tracking.There are fourteen active cells in three image sequences which have been predicted andtracked in the paper simulation experiment. The experimental results show that the errors ofactive cells estimated values and observation values in three image sequences are less than2.5pixels, so the EKF-PF algorithm may approach to the solution of accurately and effectivelyprediction and tracking of the active cells.更多还原