Härtill 8 uppsatser

Diss. (sammanfattning) Helsingfors : Aalto-universitetet, 2025

The aim of the thesis is to investigate state-of-the-art methods for isolated speech recognition using spiking neural networks, non-contact respiratory monitoring using thermal imaging, and multi-sensory approaches for ship and sea ice detection. The applications include methods for sound based ship foghorn bearing estimation, bearing estimation in visible and thermal infrared imaging, sea ice-track semantic detection for navigation, and generic multi object tracking. Although these applications span diverse domains, they are unified by the core challenge of extracting robust, structured information from noisy, time-varying sensory data. We explore architecture of liquid state machines to classify speech patterns. Using spiking neurons and models of cochlear processing, we introduce a novel performance measure, named memory metric to evaluate system's ability to classify speech. We non-invasively analyze breathing patterns by extracting airflow signals from nasal temperature changes captured with a thermal camera. Auditory and thermal signals are both temporally processed. In maritime awareness, we use sound sensors to estimate ship bearing, and visible and thermal infrared vision sensors are combined with deep learning and Kalman filters for multi-modal ship tracking. For navigating icy regions, we apply deep learning to identify sea ice in images, using fused visible and thermal imagery. Together, these domains highlight how data fusion, tracking, and classification can be applied across vision, acoustics, and thermal sensing.