Non-speech Acoustic Event Detection Using Multimodal Information

Non-speech acoustic event detection (AED) aims to recognize events that are relevant to human activities associated with audio information. Much previous research has been focused on restricted highlight events, and highly relied on ad-hoc detectors for these events. This thesis focuses on using multimodal data in order to make non-speech acoustic event detection and classification tasks more robust, requiring no expensive annotation. To be specific, the thesis emphasizes designing suitable feature representations for different modalities and fusing the information properly. Two cases are studied in this thesis: (1) Acoustic event detection in a meeting room scenario using single-microphone audio cues and single-camera visual cues. Non-speech event cues often exist in both audio and vision, but not necessarily in a synchronized fashion. We jointly model audio and visual cues in order to improve event detection using multistream HMMs and coupled HMMs (CHMM). Spatial pyramid histograms based on the optical flow are proposed as a generalizable visual representation that does not require training on labeled video data. In a multimedia meeting room non-speech event detection task, the proposed methods outperform previously reported systems leveraging ad-hoc visual object detectors and sound localization information obtained from multiple microphones. (2) Multimodal feature representation for person detection at border crossings. Based on phenomenology of the differences between humans and four-legged animals, we propose using enhanced autocorrelation pattern for feature extraction for seismic sensors, and an exemplar selection framework for acoustic sensors. We also propose using temporal pattens from ultrasonic sensors. We perform decision and feature fusion to combine the information from all three modalities. From experimental results, we show that our proposed methods improve the robustness of the system