A joint Feature Extraction and Data Compression Method For Low Bit Rate Transmission In Distributed Acoustic Sensor Environments

Unattended passive acoustic sensors are among the widely used sensors for remote battlefield surveillance, situation awareness and monitoring applications. These small and cost effective sensors can provide real-time information about different types of ground and airborne targets. They are rugged and reliable and can be left in the field for a long period of time after deployment. To improve the spatial resolution for separating multiple closely spaced targets that move in tight formations while reducing the on-board computational requirements, a modest quantity of single microphones could be deployed in a surveillance area of interest. These microphones are considerably less expensive, small sized and contain generic DSP boards capable of performing simple detection, feature extraction and data compression tasks. They are also equipped with basic communication systems to transmit essential compressed target information to a master station which has more sophisticated computational power to carry out high-level operations for sensor array processing and target detection, tracking, and classification. This Phase I research involves development of a joint feature extraction-data compression/encoding method for low bit rate transmission of essential target information to a master computer. The extracted subband features would allow for detection and preliminary classification of the targets in time-frequency as well as data compression and encoding. In this framework, only essential frequency and tonal target features that are needed for accurate target localization and classification will be encoded and transmitted, thus yielding a low bit rate without incurring degradation in the overall detection, tracking and classification performance. This study will also propose new methods to adaptively form sensory arrays based upon coherence information. The effectiveness of the developed schemes will be demonstrated on real and synthesized data sets.