Projects

Designed and developed the Panorama application (including UI design) for Android Camera in Android Ice Cream Sandwich 4.0, and worked closely on its deployment with the Google Android Camera Team [App Source Code][Algorithm Source Code]. Designed and developed Video Stabilization technology for Google Talk on Android Honeycomb 3.0+ Tablets and facilitated deployment onsite at Google.

I led and developed core perception algorithms for a vision-based autonomous ground vehicle as part of Sarnoff’s DARPA Urban Challenge team – Team Autonomous Solutions.

Overview Talk by Mike Geertsen, Program Manager for the DARPA VMR Program. The DARPA Visual Media Reasoning Program [DoD News] was a multi-year [2012-2016], multi-million dollar project with an objective to create visual exploitation and indexing tools to rapidly extract mission-relevant visual intelligence from large quantities of diverse, ill-defined, unstructured imagery captured from multiple adversary sources. I was the technical lead on the large-scale image clustering effort and developed novel algorithms to perform unsupervised clustering on millions of unstructured images and videos into inherent semantic groups that allowed fast discovery and search.

I pursued a full-time PhD in parallel with my full-time research position as a Principal Research Scientist at SRI International, and finished my PhD in 5 years (2010-2014). Developed novel algorithms for image matching under disparate appearance (e.g. matching between image pairs like day-night, historic-current, painting-picture) using spectral correspondence. Developed purely geometric algorithms for geo-localization of street-level imagery using reference digital elevation data. Developed novel technology to match between aerial-oblique and street-view imagery of building facades with applications to image-based geo-localization.

Photos and videos taken from cell phones, laptop cameras, and other devices often contain critical information about the locations of the people, buildings, and objects seen in the imagery. The widespread availability of these images and videos has greatly expanded the volume of information requiring labor-intensive analyst review, even though quick assessment may be necessary for certain operations. Because of the lack of geolocalization tools for metadata-free images and videos, analysts must manually geolocate these images and videos by visually comparing them with terrain models and overhead images.

Designed system and algorithms for detecting pedestrians at a high detection rate and low false-positive rate by combining contextual cues from buildings, vehicles and other non-pedestrian classes as part of this project from the Federal Highway Administration (FHWA) Exploratory Advanced Research (EAR) Program (FHWA-HRT-11-056). The research results are included in the FHWA EAR summary.

Grant Page Abstract Glaucoma diagnosis, management and research depend on complex judgments of the optic disc (or optic nerve head), visual field and intraocular pressure. The current standard of optic disc evaluation requires qualitative observer judgments of stereoscopic photographs of the optic disc, a less than optimal method. Despite much research, no methods have yet conclusively improved over this conventional Approach: Contemporary optic disc analyzers typically use instrument-specific image capture methods and derive quantitative estimates for various anatomical features of the optic disc.

Formulated novel algorithms and demonstrated a full working system for semantic identification of indoor structures like doorways, stairs etc. from a scanning LIDAR equipped robotic platform as part of this program from the Robotics Technology Consortium, Inc. (RTC). Designed and developed a novel streaming architecture for real-time data processing from mobile robots.

Developed Pedestrian Tracking from vehicle-mounted IR cameras that has been deployed on the BMW 7-Series. Developed Stereo-based Adaptive Cruise Control from a vehicle-mounted camera that performs real-time detection, tracking and range-estimation of vehicles ahead of the host vehicle. Designed a stereo analysis framework that allows empirical evaluation of any given stereo algorithm. The evaluation results can be used to choose appropriate stereo rig design parameters thus allowing for use of stereo vision sensors in practical automotive environments.