Biological Motion Analysis
From sub-cellular organelles to complex organisms, many problems in biological research involve the motion analysis of multiple objects in video. We develop algorithms for the rapid mining of motion patterns in large biological video data sets.
Contact: Shin, Souvenir
We provide an interactive, clinical learning environment complete with patients in specific scenarios that are able to generate context-aware responses.
Robust Medical Image Segmentation
Organ shape plays an important role in various clinical practices, e.g., diagnosis, surgical planning and treatment evaluation. Robust segmentation of organ shapes is a fundamental problem in medical imaging. We use a learning-based deformation model and robust shape priors inspired by compressed sensing. A sparse set of shapes in the shape repository is selected and composed together to infer/refine an input shape, which is formulated as a sparse learning problem. The a priori information is thus implicitly incorporated on-the-fly.
Cardiac Ultrasound Analysis
Due to poor transmission caused by air in the body cavities after trauma or surgery, the ultrasound images from critical-care patients are much noisier than in the typical case, and automated methods designed for “clean” images fail. We develop algorithms that deal with the additional challenges of ultrasound images from critical-care settings.
Camera Networks for Activity Analysis
Can camera networks be used to understand how people use spaces over time? Can we encode days, months, or years of human activity from video? Can human behavior experts intuitively analyze this data? Our goal is to understand how public spaces are used over time using a network of ceiling mounted cameras by developing computer vision and machine learning algorithms. The specific tasks include:
- Tracking: Detecting and following people (and their gaze direction) as they move through the space
- Action Recognition: Identifying common actions, such as waving, pointing, sitting, standing, etc.
- Distributed Computing: Coordinating these algorithms over multiple cameras in the network
- Privacy Preservation: Encoding activity data without identifying people or saving video data
Large-Scale Medical Image Retrieval
As the amount of medical image data increases, there is a need for large-scale, content-based image retrieval. Applications include: image-guided diagnosis, clinical decision support, education, and efficient data management. Our methods use scalable indexing and searching algorithms for real-time image retrieval among massive and high-dimensional databases. The retrieved images indicate the most likely diagnosis, whichcan be used for clinical decision support.