Research On Multi-Cell Recognition And Tracking Technology In Video Image Sequence

In recent years,computer vision technology has developed rapidly in many fields which include vehicle detection and tracking in traffic monitoring,face recognition in intelligent security,cell behavior analysis in biomedical field and so on.It is siginificance for human disease diagnosis and biomedical research which about cell behavior analysis in biomedical field.Thus the automatic recognition and tracking methods of multi-cell have attracted extensive attension.Existing methods are mainly focus on general object,but cells are different from other objects in that their behavior includes not only birth,migration,deformation and death,but also mitosis,which makes the general multi-target recognition and tracking methods incapable for cells.In order to solve the quantitative problems of multi-cell including deformation,agglomeration,division and adhesion,this thesis takes multi-cell in microscopic video image sequences as major research object and corresponding works are as follow: 1.Multi-cell segmentation and mitosis prediction based on deep learning;2.Multi-cell automatic tracking technology based on combination of detection and mitosis prediction;3.A heuristic method for labeled ant colony track-to-track association.Corresponding innovation work and main research contents are as follows:1.Considering the spatio-temporal correlation of cell movements in image sequence,convolutional long short-term memory network is introduced into U-net segmentation network so that the function of U-net is expanded from semantic segmentation to multi-classification segmentation including mitosis prediction.Experimental results show that the proposeded network can simultaneously achieve multi-cell segmentation and mitosis prediction,which provides an experimental basis for the following research of multi-cell tracking technology.2.With the obtained detections and mitotic probability of cells,multi-cell antomatic tracking methods based on detection and mitosis prediction are studied in this paper.Particle filter realizes recursive Bayesian filtering through non-parameterized Monte-Carlo simulation.In this paper,a particle filter is initialized in the first frame for each cell for tracking.Data association is simplified into two steps: optimal assignment using Hungarian algorithm and the second assignment for mitosis.According to the predicted cell mitotic probability,an improved joint tracking approach for cell analysis is proposed in which the labeled multi-bernoulli random finite ant colony is defined,and the matching between cells and detections are optimized by the objective function of OSPA-T and achieved by the decision-making behavior of labeled ant colony.Cell lineage trees,cell states and morphological estimation are finally obtained through trail and food pheromone fields.Experimental results show that the two proposed automatic tracking methods can be applied to multi-cell tracking scenarios such as agglomeration and division.3.Track breakage and loss often occur due to missed detection,large difference in cell dynamics,mitosis,etc.,which affects the track quality of multi-cell tracking.We present a novel cell track-to-track association approach that rebuilds lineage trees through the pheromone field of a proposed ant colony optimization.With the constraint of maximum inter-frame displacement,the algorithm can link potential tracks by minimizing the proposed cost function considering both cell motion and morphology that mainly occurs on the fragmented intervals.Two different decisions are defined for ant colonies to predict mitotic and non-mitotic events used to construct relevant trail pheromone fields.A novel subsequent processing technique including threshold processing,trail merging and identity fusion is ultimately proposed,that makes full use of the spatial information of the pheromone field to realize track-to-track association.Experimental results show that the proposed algorithm can recover the track fragments and reconstruct complete cell movement tracks.There are 54 figures,5 tables and 91 references in this thesis.