CPS: Synergy: Distributed Sensing, Learning, and Control in Dynamic Environments

Abstract:

This poster presents progress of the synergistic framework and algorithms development for a group sensors to collaborate in disaster scenarios. A number of fundamental research results were obtained on scene understanding in a network of visual sensors. These were experimentally evaluated, and datasets and results from these evaluations have been released to the community. Fundamental theory and algorithmic research was performed and experimental validation was carried out in the areas of distributed estimation and planning for spatio-temporal processes, distributed people tracking from fixed and mobile robots, and optimal planning for uncertain processes such as exploration. In addition, areas related with data and uncertainty were investigated. A new idea was introduced for recognition of humans in a network of non-overlapping cameras where instead of directly matching people by their appearance, the matching is conducted in reference space where the descriptor for a person is translated from the original color or texture descriptors to similarity measures between this person and the exemplars in the reference set. This idea was extended from a pair of cameras to a network of cameras. A multi-camera tracking system with integrated crowd simulation is proposed in order to explore the possibility to make homography information more helpful. Two crowd simulators with different simulation strategies are used to investigate the influence of the simulation strategy on the final tracking performance. The experimental results demonstrate that crowd simulators boost the tracking performance significantly, especially for crowded scenes with higher density. Most recognition methods require manually labeled data in order to build the models based on which the recognition is done. One of the questions we have addressed is how such labeling can be done online as more and more data is available in a multi-sensor environment? The combination of multiple sensors and online labeling sets this work apart from many methods in the existing literature. Distributed estimation and planning algorithms are developed for complex processes such as gas leaks and forest fires. A complex spatial-temporal problem (e. g. gas leak on Cornell's campus) is modeled, and the source location estimation problem is studied. Then simplified approaches to optimal robot path planning are also studied to minimize time to estimate the source location. We explored ways of using existing uncertain databases, such as Orion, Trio and MayBMS, both with respect to uncertainty management as well as query evaluation. We also explored the querying on uncertain and probabilistic graphs. Our approach of using the deterministic graph D obtained from the probabilistic graph P to answer such queries has shown promising results. We have already applied this approach to the problem of shortest paths, and are currently extending it to the problems of determining the top-k results and to outlier detection.