Image segmentation and object tracking are two fundamental problems in computer vision and have been studied for decades. When applied to biomedical images, general computer vision algorithms for segmentation or tracking may not achieve satisfactory results, due to special characteristics of biomedical images (e.g., highly anisotropic dimensions of 3D biomedical images). In this dissertation, we will introduce new deep learning based segmentation algorithms for 2D and 3D biomedical images, and new tracking algorithms for identifying the motion of bacterial cells as well as other related applications (i.e., visible feature based iris recognition). Our new 2D deep learning model explores multiscale feature reuse to tackle extremely challenging segmentation tasks. Our 3D model employs a new paradigm to explicitly handle and leverage the anisotropism of 3D biomedical images. For the cell tracking problems, we propose a new matching model based on Earth Mover's Distance and develop a group of approaches based on this new EMD matching model for tracking different types of cells. Evaluated by biomedical images from real experiments, our algorithms can obtain segmentation and tracking results of remarkable quality and show high robustness in practice.