|M.Sc Student||Zeev Kaplan|
|Subject||Dynamic Tracking with Geometric Active Contours|
|Department||Department of Electrical Engineering||Supervisor||Mr. Tannenbaum Allen Robert|
|Full Thesis text|
This work focuses on the visual dynamic shape tracking of objects as they appear in video sequences obtained by imaging sensors. We formulate the object tracking problem as a series of recognition and detection procedures of the desired object boundaries in the given image sequence. By exploiting the “coherence of motion” and other assumptions, we have managed to construct an adequate dynamic model that holds current system state. Temporal evolution of this model, based on prior knowledge and observational data acquired from the images, enables us to essentially reduce complexity and the amount of computations in detecting and recognizing the desired objects. Segmentation is a process of recognizing different features in the image and separating the image plane into regions having some kind of unique characteristic. Segmentation in our work was implemented via geometric active contours, in the implicit level-set setting. Methods of statistical shape analysis, linear PCA and LLE appeared in context of providing a shape prior and constraining the space of possible segmentation results. The task of information fusion, both prior and observational is performed in our work by using Kalman filter for rigid objects tracking and particle filter for deforming objects. All the aforementioned techniques were numerically implemented and integrated into a united algorithmic framework.
Rigid motion estimation by Kalman filtering and tracking rigid objects with active contour was demonstrated and limitations of the proposed method were discussed. The tracking algorithm based on particle filtering and active contours was developed to deal with non-linear global motion models and local deformations of the object. We have proposed an augmented state vector to include in addition to shape, global motion parameters and photometric properties also some basic geometric properties, such as object area and boundary length. The filtering step was improved by using gradient smoothing procedure during curve evolution. This technique could be viewed as approximation to performing Sobolev gradient descent, but without change of original metric on the space of closed curves which resulted in easier implementation, making it more feasible for real-time applications. Improved global motion tracking performance was achieved by using proposed importance sampling for global motion parameters based on optical flow estimation. The resulting system seems to improve the robustness of tracking in noisy and non-stationary scenes and successfully managed the tracking of both single object and multiple object scenarios with occlusions. Our algorithm was simulated on several interesting scenarios providing encouraging results.