Studies on Sensor Integration Based on Signal Level Correlation
徹志 池田 · OUKA (Osaka University Knowledge Archive) (Osaka University) · 2014
For artificial systems that behave in response to the external world, we must use sensors that observe its state.Many kinds of sensors have been developed to describe and understand outside scenes and objects, and much work has been conducted on signal processing and pattern recognition.Integrating many kinds of sensors and generating descriptions of scenes and objects are fundamental problems in this research area.The objective of this research is to propose a new sensor integration method based on signal correlation, which appears in multimodal observations using different kinds of sensors.Since previous methods associate observations in a common position coordinate, applying them to sensors that do not directly measure positions is difficult.To associate the observations in different kinds of sensors, we focus on signal correlation, which is a signal structure that appears when localized multiple sensors observe a common scene.Although the observed physical quantity is different, the changes in the scene result in correlated changes in the observed signals.By evaluating the signal correlation among multimodal observations, our method can associate observations without measuring a representation in a common coordinate and be applied to integrate various sensors.Since our method focuses on signals at the lower level of abstraction before computing the higher level features and performing pattern recognition, we call it signal level integration.The relationship among sensory signals is also important in the area of media conversion that converts signals from one modality to another.In situations when we canft use specific modality for communication and presentation, it is effective to use a different modality in a complementary manner by media conversion.In previous media conversion methods, signals in different media are associated and converted in a common symbolic coordinate such as recognized patterns and words that describe the impression of signals.However, much information is lost when we represent signals in one medium in symbols.By converting signals in one medium into another by keeping their signal correlation, many features in the original medium are converted into another medium.However, simply computing the correlation function does not extract clear and stable relationships among multimodal observations.When objects move in a scene, it is difficult to keep their multimodal correlations since each sensor only observes its local area.When integrating observations in binary representations, we need to design a suitable method to compute the correlation.Since the observations are not always stable, we must consider the instability in computing the correlations.To achieve this goal, we expand the previous signal level integration method in three points.First, we expand it so that it associates observations when the observed target moves and the correlation among sensory signals is not stable by proposing a method that estimates the target positions and simultaneously associates observations based on the maximization of the correlation among the sensory signals.Second, we propose a new signal association method for binary observations based on a statistical test.Third, when the observation confidence changes based on the situation and affects the signal correlation, it is difficult to stably associate observations.We propose an association method that evaluates observation confidence and apply it to associate the leg motion of pedestrians and wearable accelerometers to estimate stable signal correlation.Finally, we propose a new media conversion method that converts omni-directional video to sound that keeps the impressions to signals in the original media by considering the signal correlation. List of Tables 2.1Parameters used in the experiment . . . . . . . . . . . . . . . . . . . . .