A hybrid technique for tracking network structured multiple deformable objects

In this thesis, a novel hybrid approach for tracking variable numbers of network structured deformable objects is presented. The hybrid technique developed is a combination of the Network Snakes parametric active contours technique, and the Reversible Jump Markov Chain Monte Carlo (RJMCMC)-based particle filter approach. Additionally, a novel method for (semi-)automatic initialization of the network snakes is implemented. The proposed technique is applied to the real biological problem of tissue-level segmentation, and automatic tracking, of a network of cells in confocal images showing the roots of the model plant Arabidopsis thaliana. The Network Snake component is used to model the structure of cells in Arabidopsis roots, which are clustered together and delineated by shared object boundaries forming a network topology.The RJMCMC tracker is allowed to track the network node points over image sequences, and these tracked nodes are then used to control and reparameterise the topology of the network snakes at each time step. This is followed by energy minimization of the network snakes which refines the tracked nodes and cell boundaries to settle at the energy minimum. Thus the component techniques complement each other in the hybrid approach. A novel method of evaluating such network-structured multi-target tracking is also presented in this thesis, and is used to evaluate the developed tracking framework for accuracy and robustness using several real and synthetic time-varying and depth varying(z-stack) image sequences of growing Arabidopsis roots.