Laboratory for Perception Action and Cognition (LPAC)
Directors : Professors Robert T. Collins and Yanxi Liu
|
The Laboratory for Perception, Action and Cognition (LPAC) has a wide range of research topics, including low-level motion perception, real-time control of active sensors, design of multi-sensor surveillance networks, analysis of human body motion, recognition of activities and events, and mathematical group theory-based modeling of machine and human texture perception for image understanding and manipulation, dynamic near-regular texture tracking, and texture-based localization. Current application areas include moving object detection and tracking from unmanned air vehicles, recognition of human identity and action within smart spaces, real-time stereo-motion analysis for autonomous navigation, quantified 3D/4D facial asymmetry for gender/expression classification, computer aided diagnosis, large biomedical image database indexing and retrieval, and analysis and synthesis of active crowds or near-regular textures on deformable media such as cloth or through transparent fluid.
The name of the lab is motivated by the multiple disciplines that we are bringing together to develop robust intelligent systems:
|
Perception = Computer Vision
Aristotle defined Vision as the act of knowing what is where by looking. Likewise, the goal of Computer Vision is to interpret visual sensor data, including still imagery and video, to measure 3D scene structure, infer the identity and locations of objects, and recognize dynamically occurring activities and events.
|
Action = Robotics
By coupling computer vision input devices with robotic actuators that can change or manipulate properties of the real or virtual world, we are able to build feedback loops that exhibit intriguing, emergent behaviors. We take a broad view of the term actuator to include unmanned vehicles, pan/tilt/zoom heads that actively change the camera viewpoint (e.g. active vision systems), and monitor/projector systems that display useful information into or even on to the scene (e.g. smart spaces).
|
|
|
Cognition = Artificial Intelligence We believe the performance requirements of competent intelligence systems are too demanding to be programmed completely by hand, and focus instead on developing systems that can learn from training examples and from their own unsupervised exploration. We also seek to develop mathematical representations that can leverage fundamental constraints present in the physical world, be it the importance of gravity in interpreting natural imagery, the periodic nature of locomotion (e.g. gaits), the symmetry of sptiotmporal patterns or the interplay of randomness and regularity in the appearance of natural and man-made scenes.
|
LPAC Meeting & Seminar
