Improving cell tracking by jointly handling missed detections, false detections, cell divisions and higher-order motion models

在从发育生物学到细胞生物学的各个领域中，为了从细胞轨迹的定量分析中获得基本的洞察力，需要非常高的细胞跟踪精度。然而，任何现有的小区跟踪算法都无法达到所需的精度。在这个项目中，我们计划增加所谓的按赋值跟踪模型的表达能力，同时减少它们的运行时间。为了实现前者，我们将引入时间顺序更高的因素来建模先验知识，如恒定速度、轨迹的平稳性和与细胞周期的一致性。这些发展需要与近似组合优化的改进紧密结合在一起。总而言之，这些发展应该会扩大计算机视觉的最先进水平，从而最终允许在挑战生命科学的2D+时间和3D+时间数据集的情况下进行可靠的细胞跟踪。

In fields ranging from developmental to cell biology, very high cell tracking accuracy is required to gain fundamental insight from the quantitative analysis of cell trajectories. However, the required precision is not yet achieved by any existing cell tracking algorithm. In this project, we plan to increase the expressiveness of so-called tracking-by-assignment models while decreasing their runtime. To achieve the former, we will introduce factors that are higher order in time to model prior knowledge such as constant velocity, smoothness of trajectories and consistency with respect to the cell cycle. These developments need to be tightly coupled with improvements in approximate combinatorial optimization. Together, these developments should expand the state of the art in computer vision so as to finally allow a reliable cell tracking in challenging 2D+time and 3D+time datasets from the life sciences.