Summary of PhD Work (publication page more updated...)
Image based localization
Robust techniques for motion estimation and image registration
Building recognition/detection
Geometric information extraction
Building environment model from images/video
The problem of construction of visually realistic models of the surrounding environments is actively being pursued by various researchers in computer vision and computer graphics.
Exploring single view/multi-view geometric information
Efficient Vanishing Point extraction
Automatic weak model extraction and camera calibration
Given the fact that constructing realistic 3D model is difficult, a collection of images can be used to render the scene, in stead of building 3D model of the environment. Despite its limitation in generating novel views, image mosaics have the advantage of easier acquisition, thus it's a good complement to the geometric model.
Tracking based mosaic, which is built from video sequence based on point tracker.
360 panorama for our lab and hallway, built from 12 images each.
Detection/recognition of buildings
In urban environments, man-made structures like buildings are often used as landmark for localization and navigation. Motivated by this, detection and recognition of buildings in the view are investigated in this project, furthermore, relative position and orientation (camera pose) are recovered.
One of the central issues pertinent to recognition is the choice of suitable representation. Both accuracy and efficiency need to be considered when implementing system.
In an initial experiment, Scale Invariant Feature Transform (SIFT) is utilized as feature detector /descriptor for the problem. On a database of 68 buildings we collected (available as two zip files: buildinga and buildingb), 75% recognition rate was achieved. More detail..
Given the fact that color images are usually available now, a Hierarchical Building Recognition scheme is proposed to tackle the problem. The core of the scheme is two localized hue histograms which provide high discriminating power while being very compact. The whole system achieved 96.5% recognition rate on a public downloadable database ZUBUD, which is better than other previously reported results.
Vision guided robot navigation and control
The role and representation of visual information (appearance/geometric) for navigation was studied here.
One preliminary experiment I tried was vanishing point based visual servo along corridors, where geometric information like lines and points are used for indoor robot navigation.
*This research
was supported by National Science Foundation grant NSF IIS-01183232.
