| Start Date: | 5/21/2014 | Start Time: | 2:00 PM |
| End Date: | 5/21/2014 | End Time: | 3:30 PM |
|
Event Description
Title: 3D pedestrian tracking and 3D surface alignment of ceramic fragments using geometric and color cues
Advisor: Dr. Fernand Cohen
Date: Wednesday, May 21, 2014
Time: 2:00 p.m.
Location: Biomed Seminar Room 709, 7th Floor, Bossone Research Enterprise Center
Abstract
With the prevalence of video surveillance, face tracking and recognition in a surveillance system is drawing more attention and is pursued more rigorously. A high quality face image for face recognition is usually captured using a pan-tilt-zoom (PTZ) camera guided by the 3D head position of a person which is obtained from a pedestrian tracking system. Side view cameras are usually used to track persons, while using overhead cameras in indoor environment, such as train stations, airports, hotel lobbies, is understudied. Some researchers use one overhead camera and estimate the 3D head position based on the ground position and an assumption on the person’s height. Some others use a stereo overhead camera to construct a full 3D depth/disparity map for the scene. I propose to track pedestrians based on the detection of the 3D head points using 2 horizontally aligned overhead cameras. The research proposal addresses the issue of inaccuracy in estimating 3D head position when using one overhead camera and the issue of inefficiency when using a full disparity map. A short segment passing through the head top of a person is established in a frame from the left camera based on projection geometry. With the synchronized frame from the right camera, the 3D head point is detected as the highest point on the segment without the need of a full disparity map. The color information within the head top area is also integrated in head detection and tracking. An approach to schedule multiple PTZ cameras to capture a series of best quality frontal face image is also proposed based on probability measure.
The geometric and color information utilized in 2D images can be also used in 3D surface alignment which assists in the tedious task of manually reconstructing ceramic vessels from unearthed shards. Generic models are generated by experts as a rendition of what the original vessel may have looked like. 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 regions with color markings on 3D surfaces are extracted both in ceramic fragments and generic models. Ceramic fragments are then aligned against the corresponding generic models based on the geometric relation between the extracted regions. |
|
Location:
Biomed Seminar Room 709, 7th Floor, Bossone Research Enterprise Center |
Audience:
Current StudentsFacultyStaffGraduate Students |
|
