数码立体显微视频多特征点人工智能跟踪研究  

Research on Multi Feature Point Artificial Intelligence Tracking of Digital Stereo Micro Video

作　　者：陈斯宇 王培培[1] CHEN Si-yu;WANG Pei-pei(Educational Technology Center,Jilin University,Changchun Jilin 130000,China)

机构地区：[1]吉林大学教育技术中心,吉林长春130000

出　　处：《计算机仿真》2021年第4期348-351,446,共5页Computer Simulation

摘　　要：在对数码立体显微视频进行多特征点跟踪处理时,由于显微构件的倾斜旋转,以及视频目标重叠交叉等原因,导致特征点轨迹跟踪效果不理想。为此提出了基于粒子滤波状态观测的多特征点跟踪方法。首先建立基于空间变换的特征状态方程,并进一步设计了关于特征点位置、速度和噪声的观测模型,同时针对采样不均匀引入补偿处理。然后在计算状态方程的过程中,考虑到特征分布情况,设计了粒子滤波方法。利用状态矢量和对应权重构建粒子,通过粒子的分布概率更新其权重,同时根据权重计算出特征观测值。最后把特征点的动态变化和时间建立关联,将上述过程转换为限定条件下关于状态关联的寻优处理。通过关联门限和位置约束,预测得到关联状态构成的时间序列,即为特征点轨迹。通过仿真,验证了建立的多特征点状态观测模型能够准确描述显微视频中特征点的运动状态,获得更加精准稳定的跟踪性能。In the process of multi feature point tracking of digital stereo micro video, the tracking effect of feature points is not ideal due to the tilt rotation of the micro components and overlapping of video targets. Therefore, a multi feature point tracking method is proposed based on particle filter state observation. Firstly, the characteristic state equation based on the spatial transformation was established, and the observation model about the position, velocity and noise of the characteristic points was designed. At the same time, compensation processing was introduced for the nonuniformity of sampling. And then in the process of calculating the equation of state, considering the characteristic distribution, the particle filter method of PHD was designed. The state vector and corresponding weight were used to construct particles, and the weight was updated by the distribution probability of particles. At the same time, the characteristic observation value was calculated according to the weight. Finally, the dynamic changes of feature points were correlated with time, and the process was transformed into the optimization of state correlation under limited conditions. Through the correlation threshold and position constraints, the time series composed of the correlation state was predicted as the characteristic point trajectory. Through the simulation experiment, it was verified that the multi feature point state observation model can accurately describe the motion state of the feature points in the micro video, and obtain more accurate and stable tracking performance.

关 键 词：数码立体显微视频 多特征点跟踪 空间变换 粒子滤波 状态观测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]