Computational Sensing and Perception
Speaker: Joshua Smith , Escher Labs
Date: April 3 2003
Time: 4:00pm
Location: NE43-941
Machine perception research often aims to create machine versions of human senses, and replicate human capabilities. This talk describes three machine perceptual channels that are inaccessible to human perception. In some cases these sensory modalities are intrinsically valuable, but have been overlooked because they are not shared with humans; in others, the value is explicitly tied to the fact that they are imperceptible.
Electric Field Imaging is a channel for machine sensing and perception that has been overlooked. Like Computer Vision, Electric Field Imaging (EFI) aims to extract useful spatial information from sensor data; unlike vision, it relies on low frequency electric fields rather than visible light. I will describe physical mechanisms, instrumentation, signal processing, and inference algorithms for EFI. Specifically, I will present an Electric Field Imaging technique that is analogous to the photometric stereo method in computer vision: the region of interest is "illuminated" with several distinct incident fields, which induce several distinct sets of measurements ("images"). Then a computational process infers the single conductivity distribution ("scene") that best explains the data. Electric Field Imaging has lead to a number of successful practical applications, including an airbag control system that has been deployed in millions of cars worldwide. The system senses the body configuration of the passenger and is thereby able to make more intelligent firing decisions.
The fields used in EFI measurements are modulated, which allows a rich set of concepts from communication theory to be applied to sensing. I will explain how a software implementation of code division multiplexed Electric Field Imaging suggests a notion of generalized "foveal" and "peripheral" perception for network sensing systems, and points to the possibility of co-design of perceptual and sensory computational processes in such systems.
The modulation techniques used in EFI can be applied in very different settings to enable other machine perceptual capabilities. I will explain how techniques used for EFI measurements lead me to develop steganographic algorithms and channel models. In the case of steganography, the value of the channel is explicitly tied to its imperceptibility. Finally, I will describe FiberFingerprint Identification, a security technology which allows a machine to uniquely identify and track items that are indistinguishable to human perceptual processes.
See other events that are part of AI Colloquium Series Spring 2003
See other events happening in April 2003
