United Technologies Research Center (UTRC) is the innovation engine and research vehicle for United Technologies Corporation (UTC), and serves to solve challenging problems in perception, robotics and controls technologies for its business units such as OTIS, Pratt & Whitney, Climate Control and Security, and Aerospace systems. UTRC also works with government on various DOD, DARPA and ARM funded research.
In this talk, I will present the work done in the area of human action and activity localization from streaming videos. Here, non-linear manifolds and the grammar/codewords are learned using auto-encoders and conditional restricted Boltzmann machines for each category of action. For inference, these learned manifolds are traversed by the features of the test video segment to get action class and its percentage of completion at each frame. This work provides a way to realize real time human action localization with possibility of predicting the next action or sub-action from a short streaming segment of frames invariant to the speed of motion of action and frame rate of camera. Based on this work, I will discuss some of the research efforts and next steps that UTRC focusses on towards realizing human robot collaboration and human aware navigation for improving manufacturing outcomes in assembly operation in unconstrained environment.
Binu Nair is a Senior Research Scientist with United Technologies Research Center at Berkeley where he focusses on computer vision and deep learning algorithms for next-gen robotic perception and automation systems. His research interests include object tracking, person identification, and activity recognition with emphasis on human-aware robot navigation, and human robot interaction. Prior to this work, he was a Research Engineer with University of Dayton Research Institute where he built novel deep learning algorithms for machine part feature detection and recognition to automate human-level inspection tasks in manufacturing. Binu graduated with a PhD in Electrical Engineering from University of Dayton in 2015, where the dissertation was on human action recognition and localization from streaming videos. He has published for 15+ articles in top publications and is a reviewer for IEEE Transactions in Image Processing (TIP) and Journal of Electronic Imaging (JEI). Binu is also passionate about promoting diversity in STEM and has had the opportunity to drive this mission by presenting in tech conferences and universities such as UC Berkeley.