Automatic Acoustic Events Detection, Classification, and Semantic Annotation for Persistent Surveillance Applications

Amjad H. I Alkilani, Tennessee State University

Abstract

Acoustic surveillance and human behavior analysis represent some of the ongoing research topics in signal processing. Acoustic sensors offer a promising sensing modality, primarily because they can capture a huge amount of information from the environment. Moreover, they can be rapidly deployed and are low-cost. In the past, significant efforts have been devoted to detecting sounds of individual objects or events. However, the issue of understanding human activities based on sporadic acoustic sound events has received unequal attention in the literature and hence is not well understood. This dissertation presents an extensive literature survey on this topic and discusses existing advanced techniques for acoustic signal processing and pattern recognition. A novel theoretic framework (Acoustic Events Detection, Classification, and Annotation (AEDCA)) is proposed which accommodates sound events ontology for improved human activities recognition. Based on a generalized taxonomy, three sound categories signifying interaction of human with each other, with vehicles, and with other objects are introduced. In order to understand different type of human interactions salient sound events are preliminarily identified and classified based on trained set of data. To interlink salient events representing an ontology-based hypothesis, a Hidden Markov Model-Acoustic Activity Recognizer (HMM-AAR) is modeled to recognize spatiotemporally correlated events. Once such a connection is established, an annotation of perceived sound activity is generated. The performance of the AEDCA system was tested and measured experimentally in both indoor and outdoor environments. Appropriate confusion matrices are developed for the assessment of performance reliability, and computational efficiency of the AEDCA system. The obtained results are very promising and strongly demonstrate the AEDCA is both reliable and effective, and can be extended to future surveillance applications.^

Subject Area

Information Technology|Engineering, System Science|Physics, Acoustics

Recommended Citation

Amjad H. I Alkilani, "Automatic Acoustic Events Detection, Classification, and Semantic Annotation for Persistent Surveillance Applications" (2014). ETD Collection for Tennessee State University. Paper AAI3623002.
http://digitalscholarship.tnstate.edu/dissertations/AAI3623002

Share

COinS