3D Pedestrian Tracking and Virtual Reconstruction of Ceramic Vessels Using Geometric and Color Cues.

摘要

Object tracking using cameras has many applications ranging from monitoring children and the elderly, to behavior analysis, entertainment, and homeland security. This thesis concentrates on the problem of tracking person(s) of interest in crowded scenes (e.g., airports, train stations, malls, etc.), rendering their locations in time and space along with high quality close-up images of the person for recognition. The tracking is achieved using a combination of overhead cameras for 3D tracking and a network of pan-tilt-zoom (PTZ) cameras to obtain close-up frontal face images. Based on projective geometry, the overhead cameras track people using salient and easily computable feature points such as head points. When the obtained head point is not accurate enough, the color information of the head tops across subsequent frames is integrated to detect and track people. To capture the best frontal face images of a target across time, a PTZ camera scheduling is proposed, where the 'best' PTZ camera is selected based on the capture quality (as close as possible to frontal view) and handoff success (response time needed by the newly selected camera to move from current to desired state) probabilities. The experiments show the 3D tracking errors are very small (less than 5 cm with 14 people crowding an area of around 4 m2) and the frontal face images are captured effectively with most of them centering in the frames.;Computational archaeology is becoming a success story of applying computational tools in the reconstruction of vessels obtained from digs, freeing the expert from hours of intensive labor in manually stitching shards into meaningful vessels. In this thesis, we concentrate on the use of geometric and color information of the fragments for 3D virtual reconstruction of broken ceramic vessels. Generic models generated by the experts as a rendition of what the original vessel may have looked like are also utilized. The generic models need not to be identical to the original vessel, but are within a geometric transformation of it in most of its parts. The markings on the 3D surfaces of fragments and generic models are extracted based on their color cues. Ceramic fragments are then aligned against the corresponding generic models based on the geometric relation between the extracted markings. The alignments yield sub-scanner resolution fitting errors.