Research

Computer Vision

My research interests in computer vision include text and human activity detection and recognition, 3D modeling/graphics with application to human face and expression analysis, semi-automatic computer vision with minimum human efforts (human computer interaction), and application of machine learning/pattern recognition in these areas. The research works in this area will have a direct impact in the real world.

I have been affiliated with several research groups in CMU. In Interactive Systems Lab in Human Computer Interaction Institute, I worked on text detection and translation, which is a foreign text/sign detection and translation project, partially supported by DARPA under the TIDES project. I developed an adaptive algorithm to detect texts from natural scenes. I finished a fully automatic, real time system for text detection and translation.

In Face Analysis group, Robotics Institute, I worked on stabilization of human face using 3D tracking. Now I am working on 3D human face modeling. Stabilization makes facial expression analysis feasible in the situation of large head motion. 3D face modeling is important for human face and expression analysis in the future. 

In HumanID (Human Identification from a Distance) group, I am now working on human upper body gesture detection, tracking, and recognition. Most previous works on human gesture tracking assume that detailed location of human body parts are given before tracking start. In our system, we use Bayesian inference and a multi-frame probabilistic algorithm to acquire this information automatically. The algorithm takes the advantage of strong spatial and temporal constraints in human motion.

My recent work is focused on a layered approach for articulated human motion segmentation. We treat articulated human motion analysis as a problem of parametric motion pattern retrieval from video. A two step process is applied: 1. Motion pattern finding; 2. Dynamic layer extraction.

Talks

Selected publications in this area. 

Image Processing, Pattern Recognition and Handwriting Recognition 

I did research works in OCR and pattern recognition in Institute of Image and Graphics, Department of Electronics Engineering, Tsinghua University, China. The main focus was on  handwritten character recognition, shape modeling/ extraction, and combined statistical and structural approaches  for pattern classification.  I taught the course: “Advanced Digital Signal Processing” ( “Statistical Signal Processing” and “Wavelets/Time-frequency Analysis”) for graduate students in Tsinghua. 

We used the concept of  image cells in shape modeling and extraction. Image cells play the similar role as phonemes in speech recognition, and are extracted using a selective attention approach. A dynamic programming algorithm was developed for  applications of the concept in handwritten character segmentation and recognition. We also proposed  new algorithms to improve the linear discriminant analysis (LDA) method for pattern classification.

Selected publications in this area. 

Statistical Modeling and Estimation

I obtained my B.S. and Ph.D. degrees in Electrical Engineering from Xi'an Jiaotong University and Northwestern Polytechnical University, both in Xi'an, China. My Ph.D. research direction was in Signal Processing and Control Theory. My research works were focused on statistical signal processing, model identification, and time-frequency analysis. 

We applied blind system identification, optimal estimation, and adaptive signal processing to the inverse problem of Seismic Signal processing. We model earth seismic systems using BSI and maximum-likelihood methods, detect and estimate major seismic events using Kalman filtering/smoothing and signal detection algorithms.        

In modeling and estimation, many of my works were focused  on frequency-domain methods , which is important for analyzing the fundamental structures and properties of systems. My works in this area include:  A frequency-domain spectrum estimation approach for adaptive Kalman filtering, a fast frequency-domain approach for signal deconvolution using fixed interval smoother, several algorithms for optimal detection of Bernoulli-Gaussian sequences  based on optimal smoothing, and estimation of model error bounds for frequency-domain identification algorithms. I also used higher order statistics for blind identification of non-minimum phase systems.

Selected publications in this area.   

Academic and Research Awards

4th place in province, in National Mathematical Contest in China.  

Graduated from the Special Class for Gifted Young at Xi’an Jiaotong University.

First Award for the Advancement of Science and Technology, Dagang Oilfield, Tianjin, China, 1994.  

Ph.D. Dissertation Award, Northwestern Polytechnic University.  

Young Faculty /Researcher Award, Tsinghua University, 1999.