This paper proposes feature extraction methods for object classification with passive acoustic sensor networks deployed in (sub-)urban environments. We analyzed the emitted acoustic signals of three object classes: guns (muzzle blast), vehicles (running piston engine) and pedestrians (several footsteps). Based on the conducted analysis, methods are developed to extract features that are related to the physical nature of the objects. In addition, localization methods are developed (e.g. pseudo-matched-filter), because the object location is required for one of the feature extraction methods. As a result, we developed a proof-of-concept system to record and extract discriminative acoustic features. The performance of the features and the final classification are assessed with real measured data of the three object classes within sub-urban environment.

Importance of this researchAlmost any process that undertakes actions relies on situational awareness. Such awareness can be supported by using sensors. In dynamic situations the observation task can get complicated and sensors may need to be reconfigured in real time. Choosing the right configuration and allocating the available resources, such as time, of the sensors is very challenging. This thesis proposes the concept of mission-driven resource management to automatically and optimally decide on such choices during the various sensor life cycle phases.

No abstract

An efficient strategy solution is developed for a specific deployment problem in which different types of sensors are required to simultaneously cover the same area of interest. The deployment goal is to select the sensor positions and orientations in such a way that the sensor network coverage is optimized. A general challenge within resource allocation problems is that, even with small-scale sensor networks, the number of possible final deployment solutions expands very fast and the problem becomes intractable. We assume that it is acceptable to trade solution optimality against algorithm speed. In this case, algorithms can be based on greedy and/or divide-and-conquer principles, which both results in good computational efficiency. We developed an efficient algorithm in three steps. Firstly, we developed a global search algorithm, but with improvements that reduce the search space significantly without losing optimality. Secondly, we transformed the global algorithm into a sequential and a hierarchical algorithm for more efficiency at the cost of optimality. Thirdly, we combine the sequential and hierarchical principles into one algorithm which results in even higher efficiency. In the end, the algorithms are evaluated with the use of an extensive testing scheme which generates many random cases for solving.

No abstract