This page contains selected research and class projects.
Vertical LiDAR Registration
The aim of this project is to reconstruct a virtual 3D environment by registering successive
2D laser range scans (LiDAR). Registration means computing the relative movement of
the LiDAR equipment as we capture data. Currently we use an additional LiDAR/stereo camera
for registration. Very soon we will extend the scenario to urban outdoor scenarios within
the UCSC campus. (current work with Suresh Lodha)
Tracking GPS-Equipped Objects Using the Experts Framework
In this work, we apply an expert learning framework for
predicting and learning the motion of an uncertain mobile
object in real-time.
We define a number of probabilistic experts, each of which predicts
the future position of the object with some uncertainty
and then combine the predictions of all the
experts to produce an estimate of the object's location. Individual
experts predictions are weighted adaptively depending on their performance.
We show that this adaptive combination is powerful
when there are changes in the pattern of the object's motion.
We have tested our algorithm with synthetic data
using uniform and non-uniform patterns as well as
real data acquired using GPS equipment in presence of intermittent and
highly erroneous data. (work with Suresh Lodha)
Collaborative filtering algorithms for the EachMovie dataset
I implemented a family of multiplicative algorithms which seem to be promising, in that they clearly out-perform
recently developed collaborative filtering
algorithms for the EachMovie dataset. The properties if these algorithms make them possible candidates for application in
other domains. (research done during independent study with David Helmbold and Manfred Warmuth in Spring'02)
Dense Depth Recovery By Dynamic Programming
I implemented a dynamic depth recovery algorithm for stereo
pairs with large occlusion. I tried to improve the results using the Mean Shift filter to smoothen images and increasing
the number of ground control points by incorporating edge features. There was no significant improvement. (course project
for Image Analysis and Computer Vision in Winter '03)
Sparse Disparity Computation via Plane Induced Homography
When there is large camera motion, it is hard to
recover robust point correspondences. However if there is a large plane in the image, we can use the homography induced
by this plane to greatly improve the feature matching results. This method is particularly useful in urban environments
where building faces can be used to induce homographies. (course project for Advanced Computer Graphics in Spring '03)
Classifying DNA Hairpin molecules by behaviour in a Nanopore
DNA strands caught in an \(\alpha\)-hemolysin
channel cause a blockade in the ionic current flowing through the channel. The current blockade waveform is
characteristic of the type of DNA molecule caught in the channel. I applied adaBoost for classification and evaluated the
``goodness'' of the features I used. (course project for Machine Learning in Fall '01)