DESIGN OPTIMIZATION OF EMBEDDED SIGNAL PROCESSING SYSTEMS FOR TARGET DETECTION
MetadataShow full item record
Sensor networks for automated detection of targets, such as pedestrians and vehicles, are highly relevant in defense and surveillance applications. For this purpose, a variety of target detection algorithms and systems using different types of sensors have been proposed in the literature. Among them, systems based on non-image sensors are of special interest in many practical deployment scenarios because of their power efficiency and low computational loads. In this thesis, we investigate low power sensor systems for detecting people and vehicles using non-image sensors such as acoustic and seismic sensors. Our investigation is focused on design optimization across trade-offs including real-time performance, energy efficiency, and target detection accuracy, which are key design evaluation metrics for this class of systems. Design and implementation of low power, embedded target detection systems can be decomposed into two major, inter-related subproblems: (a) algorithm development, which encompasses the development or selection of detection algorithms and optimization of their parameters, and (b) system development, which involves the mapping of the algorithms derived from (a) into real-time, energy efficient implementations on the targeted embedded platforms. In this thesis, we address both of these subproblems in an integrated manner. That is, we investigate novel algorithmic techniques for improvement of accuracy without excessive computational complexity, and we develop new design methodologies, tools, and implementations for efficient realization of target detection algorithms on embedded platforms. We focus specifically on target detection systems that employ acoustic and seismic sensing modalities. These selected modalities support the low power design objectives of our work. However, we envision that our developed algorithms and implementation techniques can be extended readily to other types or combinations of relevant sensing modalities. Throughout this research, we have developed prototypes of our new algorithms and design methods on embedded platforms, and we have experimented with these prototypes to demonstrate our findings, and iteratively improve upon the achieved implementation trade-offs. The main contributions of this thesis are summarized in the following. (1). Classification algorithm for acoustic and seismic signals. We have developed a new classification algorithm for discrimination among people, vehicles, and noise. The algorithm is based on a new fusion technique for acoustic and seismic signals. Our new fusion technique was evaluated through experiments using actual measured datasets, which were collected from different sensors installed in different locations and at different times of day. Our proposed classification algorithm was shown to achieve a significant reduction in the number of false alarms compared to a baseline fusion approach. (2). Joint target localization and classification framework using sensor networks. We designed a joint framework for target localization and classification using a single generalized model for non-imaging based multi- modal sensor data. For target localization, we exploited both sensor data and estimated dynamics within a local neighborhood. We validated the capabilities of our framework by using an actual multi-modal dataset, which includes ground truth GPS information (e.g., time and position) and data from co-located seismic and acoustic sensors. Experimental results showed that our framework achieves better classification accuracy compared to state of the art fusion algorithms using temporal accumulation and achieves more accurate target localizations than a baseline target localization approach. (3). Design and optimization of target detection systems on embedded platforms using dataflow methods. We developed a foundation for our system-level design research by introducing a new rapid prototyping methodology and associated software tool. Using this tool, we presented the design and implementation of a novel, multi-mode embedded signal processing system for detection of people and vehicles related to our algorithmic contributions. We applied a strategically-configured suite of single- and dual-modality signal processing techniques together with dataflow-based design optimization for energy-efficient, real-time implementation. Through experiments using a Raspberry Pi platform, we demonstrated the capability of our target detection system to provide efficient operational trade-offs among detection accuracy, energy efficiency, and processing speed. (4). Software synthesis from dataflow schedule graphs on multicore platforms. We developed new software synthesis methods and tools for design and implementation of embedded signal processing systems using dataflow schedule graphs (DSGs). DSGs provide formal representations of dataflow schedules, which encapsulate information about the assignment of computational tasks (signal processing modules) to processing resources and the ordering of tasks that are assigned to the same resource. Building on fundamental DSG modeling concepts from the literature, we developed the first algorithms and supporting software synthesis tools for mapping DSG representations into efficient multi-threaded implementations. Our tools replace ad-hoc multicore signal processing system development processes with a structured process that is rooted in dataflow formalisms and supported with a high degree of automation. We evaluated our new DSG methods and tools through a demonstration involving multi-threaded implementation of our proposed classification algorithm and associated fusion technique for acoustic/seismic signals.