Home Icon
     Home    
Research Icon
 Research
Publications Icon
Publications
 Teaching Icon
 Teaching
 Student Icon
Students & visitors		 SIP Icon
 SIP Lab
Prospective Student Icon
Prospective students

Real-time analysis of surveillance video data
Big data continues to grow exponentially, and surveillance video has become one of the largest sources. The surveillance-video data acquired by existing large networks of surveillance cameras introduce many technological challenges, including storage, transmission, and analysis. Among these, one of the most critical challenges is how to intelligently analyze and understand the visual information, e.g., for automatic video surveillance.

We are developing develop video anomaly detection methods suitable for real-time security applications, within the context of video data acquired by surveillance cameras in cities like London and Mexico City. This research involves designing efficient and low complexity visual data representations, e.g., through feature descriptors, and machine learning methods suitable for video data.
Person identification through time-invariant face recognition
Temporal-invariant face recognition allows identifying people using images of faces, despite the changes due to facial aging. In this project, we are designing feature descriptors and machine learning approaches that take into account changes in faces, due to aging, on their associated 3D shape and texture. We are particularly interested in employing features learned by Convolutional Neural Networks (CNNs).
Coding of big medical imaging data
Digitization of whole microscope specimens has recently become possible with the introduction of high-throughput slide scanners. The digitized versions of microscope glass slides, which are called virtual slides or whole-slide images (WSIs), offer several advantages over traditional glass slides in terms of portability, archival, retrieval, ease of sharing, and the ability to use computer-aided diagnostic tools.

A key challenge that prevents the potential of WSIs from being fully exploited is the very large file size, which poses heavy demands on storage and transmission resources. For example, a typical microscope specimen of 20 X 30 mm in size usually produces up to 40 GB of uncoded imaging data using a scanning resolution of 0.2-0.5 micro m. per pixel. Their very large file size also prevents WSIs from being fully integrated into Picture Archiving and Communication Systems (PACS) and the associated DICOM (Digital Imaging and Communications in Medicine) standard. Therefore, designing efficient and accurate coding methods capable of facilitating the access, transmission, and visualization of WSIs remains a challenge.

In this project, we are developing coding methods for access, transmission, and visualization of WSIs for telepathology applications. We are particularly interested in facilitating the integration of WSIs into current PACS while allowing interactive access to the data with resolution and quality scalability.
Rate-control for High-Dynamic Range (HDR) video coding
Rate control is an important tool that helps to control the bit rate of the compressed media. Many current video codecs, such as HEVC implementations, employ rate control algorithms that are based on models of the rate-distortion characteristics of the video sequence. Many of these models, such as the rate-lambda model used in HEVC, do not consider the wide range of luma values depicted in HDR content.

In this research, we are developing rate control algorithms for HDR video based on a multi-(rate-lambda) model approach. Our objective is to accurately attain a target bit rate while improving the reconstruction quality of HDR content, especially for very bright regions. Our algorithms are implemented and tested within the HEVC standard.