|
|
|
Updated: February, 2003
Welcome
|
The general objectives of the project are: (i) Shed light on many open scientific problems presented by the four main aspects of motion; (ii) Develop innovative and practical tools for the application domains; and (iii) Train HQP for an area of computing which will continue to grow in demand for many years to come. Technical Objectives of the project are: Trajectory acquisition and measurement: To track objects, it is necessary to use some model of object appearance. We propose to build models automatically from a video sequence by tracking features and solving for their 3D structure. For situations where trajectories of (multiple, similar) objects keep crossing-over each other, we propose to develop a scheme to resolve the cross-overs to give accurate trajectories. For images captured from multiple cameras/sensors, we propose to construct a common frame of reference. Trajectory representation: To understand how measured visual motion can support high-level interpretation tasks related to an object's identity, actions and intentions, we propose to develop a new representation language that is capable of supporting the recognition of key motion patterns over the range of scales (in both space and time) and viewpoints associated with the task. We also propose to develop new kinetic data structures to assist in tracking various proximity properties when the motion paths are not known a priori. Trajectory querying: To provide effective management of masses of trajectories, we focus on the development of new storage and retrieval schemes for spatio-temporal data. Effective retrieval relies on the development of an expressive query language and interface, and the design of new indexing schemes. Trajectory analysis and prediction: Given a database of typical motion trajectories, we seek to perform analysis to identify commonly occurring sub-trajectories and patterns. These may serve to predict in a short time interval what can happen in the immediate future, i.e., predicting from the last few steps of the trajectory, what the next few steps would be. Our success here can in fact enhance trajectory acquisition as stated in a), particularly for dealing with cross-overs. It has been noted that human actions often have observable motion pre-cursors which can predict an action even before its actual onset (e.g., drivers changing lanes). Through analysis of masses of data, we seek to identify these pre-cursors. |
There
is a fundamental gap between the geometric coordinates or pixels of raw
data and the patterns, relations, and abstractions that go into a
mental image or mental map. Whole industries are devoted to applying
software technology to bridge this gap in diverse application areas.
This project will exploit the extensive expertise and experience of the
team members in order to focus on shared solutions for transforming
image and geometric data into knowledge and information.