Jaka Krivic (2006) Segmentation and 3D tracking of superquadric modeled objects. PhD thesis.
In this thesis methods needed to model 3D objects, segment them from 3D data, and track them throughout image sequences are studied. First, the detection of articulated 3D objects is investigated. The envisioned system accepts range images at the input. Image segmentation is then performed to acquire superquadric descriptions of the scene. Also the objects are modeled with part level models described by superquadrics. Parts of an object form a structure, that distinguishes it from other objects. In order to exploit the structural difference between objects, we propose a method based on interpretation trees, which compares scene and model part by part giving object hypotheses. Various types of part matching constraints are introduced that compare scene parts to model parts in order to reduce the search for object instances. Also, a verification procedure is proposed that verifies that hypothesized scene parts really represent the object in question. This procedure also determines object position and part configuration, at least to some extent. Next, the object detection is introduced to the problem of 3D object tracking initialization. It provides the system with the initial object position and configuration, which are then further improved by fitting the part models directly to 3D data. The tracking phase takes advantage of the information about the object's position from previous frames to acquire the object's position efficiently.
Actions (login required)