Computer Graphics and Visualization Lab
Department of Computer Science at Purdue University

This is an old revision of the document!


Occlusion-Resistant Camera Designs: Acquiring Active Environments

Summary
In this project, we propose a family of Occlusion-Resistant Camera (ORC) designs for acquiring active environments despite the presence of moving interfering occluders. Being able to capture images in an in-use environment increases acquisition efficiency and quality without having to close-off the targeted site. Our cameras explicitly remove interfering dynamic occluders from acquired data in real-time and during live capture. Our key idea is to capture the scene at least twice from each viewpoint as the camera moves continually to sweep the scene and sample all surfaces. Our approach creates a single portable device combining the benefits of a stationary camera, which detects moving interfering occluders by image differencing, with those of a dynamic camera, which achieves scene coverage for inside-looking-out modeling.
ORC Designs: Individual cameras are represented by small (colored) squares. The camera's field-of-view is drawn using a small triangle for limited field-of-view cameras and as a circle for omni-directional cameras. Left: a full design of a sphere of follow cameras (color-coded) surrounding an omni-directional lead camera (black) produces occlusion-resistant images for all imaging directions and camera orientations. Middle: progressively simpler cameras require stricter control of the camera's orientation, providing motion-parallel or motion-perpendicular imaging. Right: both sequences converge to a minimum design of two individual cameras supporting all imaging directions but with strong camera-orientation constraints.
A Two-Camera Implementation (a.k.a. Lag Camera)

A snapshot of our prototype 2-view ORC.

At time 0, lead camera captures an image from a viewpoint. At a later time 1, follow camera captures another image at the same viewpoint. The foreground objects are different in these two images.

The images are smoothed using a Gaussian blur and subtracted from each other to produce a difference image. The absolute value difference image is thresholded to produce a binary image which is then subject to an image processing close operator to join nearby image components. This produces a motion mask. The motion mask can be used to select the background samples for rendering the static background scene.

Result using our method for unstructured lumigraph rendering and naive reconstruction without using our method. The moving occluder partially appears in the image.
Publications
People
 

projects/occlusion-resistant.1221506633.txt.gz · Last modified: 2008/09/15 15:23 by rosenpa
  [Main Page]   [Projects]   [Publications]   [People]   [Courses]   [Talks]