实验小鼠运动参数的模板匹配及粒子滤波提取方法  

An Extraction Algorithm for Motion Parameters of A Laboratory Mouse by Model Matching and Particle Filtering

作　　者：张继文[1,2,3] 梁桐 张淑平[4] 

机构地区：[1]清华大学机械工程系,北京100084 [2]清华大学精密超精密制造装备及控制北京市重点实验室,北京100084 [3]清华大学摩擦学国家重点实验室,北京100084 [4]清华大学生命科学学院,北京100084

出　　处：《自动化学报》2018年第1期25-34,共10页Acta Automatica Sinica

基　　金：国家自然科学基金(61403225);摩擦学国家重点实验室(SKLT09A03)资助~~

摘　　要：实验小鼠是一种变形体对象,现有方法难以从连续视频图像中同时提取出运动轨迹和体态细节.本文采用模板匹配和粒子滤波的目标跟踪方法求解这一问题.提出了一种几何体部件模型,在引入小鼠移动速率的基础上给出了其运动状态方程,以二值化前景像素与几何部件模型间的差异度方程为观测模型,以状态方程及相互独立的多维随机变量为运动模型,从而建立起基本粒子滤波算法.与逐帧差分识别方法的对比实验研究表明,所提出的模型与实验小鼠形体相似,能够达到视频在线提取的计算效率.新方法在强噪声干扰条件下解决了运动轨迹和体态同时精确估计,并有效避免了首尾识别混淆及虚影干扰等困境,从而为后续生物学行为分析提供依据.Laboratory mouse is a kind of deformable object. Existing methods can hardly extract motion trajectories and posture details simultaneously from those continuous recorded videos. An object tracking method based on model matching and particle filtering is adopted to solve this problem. A geometry based part model and its motion state function involving moving velocity are proposed. A model-observation difference function is established as the observation model by comparing the foreground pixels in the binary image and the geometry part model. A basic particle filter is built with this observation function and the motion state function with multi-stochastic variables which follow an independent distribution. Comparison is made between the proposed method and the classical frame-differencing method, which proves that the novel part model is analogous with a physical mouse in shape and supports real-time extracting rate and high computing efficiency. The novel method is able to estimate precisely both motion trajectories and posture states, and avoid effectively the faults of head-tail confusion and reflection disturbance. Therefore the novel method provides a trust worthy means for later behavioral analysis for biologists.

关 键 词：目标跟踪 粒子滤波 部件模型 实验小鼠 体态 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]