Research fields

 

Visual Sence Understanding and Motion Control for Autonomous Driving 2009~present
 
Our goal is to develop computational visual perception models and systems, which can perceptually process sensing information of dynamic environment via multiple  heterogeneous sensors , and improve the efficiency and perceptual quality of the perception system for a unmanned vehicle. We focus on incorporating into the system with mechanisms including selective attention, visual feedback, and structural information of spatial topology, etc. We also focus on intelligent motion control method for autonomous vehicles. Our research covers the following topics:
1) Computational framework for the perception of environment for unmanned vehicle, which has functions of selective attention and interactive fusion of multiple sensors;
2) Dynamical and collaborative computing model for sensing local dynamic scenarios;
3) Knowledge representation and reasoning of driving behaviors;
4) Real-time and highly reliable motion planning and controller for unmanned vehicle.

 

Visual Pattern Analysis for Advanced Coding, Object Discovery, and Retriveal. 2005~present

 

We focus on the problem of perceptual coding of image/video. We try to develop computational models to simulate mechanisms of human visual systems, and develop the perceptual mechanism-based models and algorithms. We aim to develop a series of models and methods for content-based coding, object discovery and retriveal, and we hope these models and methods can reduce the performance gap between human visual system and machine. More specifically, our research covers topics as follows:
1) Video parsing model based on visual cognitive mechanism;
2) Parsimonious representation for video coding;
3) Joint Source and Channel coding for robust and security video applications.
     
Visual Tracking Using Statistical Approaches 2002~present
 
Motion perception is one of essential functions of mammal vision, and building motion perception model that has a reasonable biologic basis plays a very important role in computer vision. Based on the Bayesian machine (Bayesian analysis + MCMC computation) and statistical learning, We focus on constructing nonlinear representation models for motion information in video with probabilistic graphical model. As to the task of model learning and inference, we imports the idea of 'survival of the fittest' in simulated evolution computing to design an efficient statistical evolution computing method. Furthermore, based on the motion representation model and computing method, we focus on three problems in visual tracking:
1)    Multi-target tracking;
2)    Nonlinear dynamics classifying and learning, Motion analysis and synthesis.