Multiple Key Points Tracking On Micro Sequences

Visual tracking technology is an important and valuable basic research in computer vision.Combined with micro-nano technology and testing technology, it has a wide range ofapplications in the field of micro-vision. The purpose of micro-vision is mainly to get the targets’motion path and three-dimensional(3-D) information, such as cells, micro components.Compared to the qualities of normal scale images and the microscopic images obtained understatic conditions, those microscopic images obtained under motion conditions have relativelymuch poorer qualities: the rotation process of the tilt stage leads to the frames with more imagenoise, such as illumination variability, motion blur, texture fuzzy, visual distortion and so on.Under those complex factors’ influence in the micro environment, it is difficult to trackmicro-targets accurately using the classical tracking algorithms for normal scale objects.Therefore, it is a more challenging work to track key points in long micro-sequences. This paperfinds that the projection’s affine invariance leads tracking of point templates to be a feasiblesolution, due to the fixed spatial relationship among the composed of simple fundamentalcomponents such as points, lines and planes. The main contents and contribution of thisdissertation as follows:1. A tilt rotation model system is designed to acquire targets3-D information undermicro-vision. The projection of complex micro camera system is analyzed and simulated.2. This section discusses the image features of long micro images sequences with poorqualities in details. The feature descriptors such as SIFT and DAISY descriptor, which alwaysare good at extracting corners for illumination variability, are used to extract, analyze andcalculate the single pixel key point’ information. However, the experiment shows that singlepixel is difficult to descript. Thus point templates including micro-neighborhood information canbe a feasible solution.3.The Kalman tracking algorithms are used to predict and track key points. Although the key point template with affluent gradient direction information, it is difficult to adapt the similartexture，non-linear movement and visual distortion merely using templates aligned with theimage registration method cannot realize accurate tracking. Therefore, during the key pointstracking, the points templates were detected by Probability Data Associating Filiter with theAmplitude Information (PDAF-AI) on the basis of key points position, and the rotating speedparameters of the state of next frame was predicted using Extended Kalman Filter(EKF), and thetemplates were properly updated. Experimental results show that the EKF tracking performanceis improved, but still cannot well adapted irregular trajectory and visual distortion problems.4. This paper proposes an adaptive particle filter (PF) of points tracking algorithm withaffine motion parameters, which can resolve the visual distortion, illumination variability andirregular motion estimation in complex micro-environments.The covariance with descriptor,which contains the intensity of templates, the polar coordinates, the first-and second-orderimage derivatives in Cartesian coordinates system, is used to detected and extract the trackedpoint templates, while the PF tracking algorithm with affine motion parameters is used to sampleand calculate the weights from those multiple point templates with affine geometrictransformation. Our PF algorithm and KLT feature tracker which have showed superiority totracking normal-scale rigid targets, are introduced to rigid multiple key points tracking in longmicroscopic image sequences. The method is tested on sequences in different environmentsunder challenging rotation velocity, amplied times, illumination conditions. The experimentalresults show that our PF algorithm are more precise and robust for rigid multiple key pointstracking in long micro sequences.5. The output datas of our PF tracking algorithm and KLT tracker are used to reconstruct3-D structures of micro components. The former is successful in reconstruction in spite of someof its key points are severe and persistent occlusions.