In this paper we propose a two-step-algorithm for the blind separation of convolutive mixtures. We show that its application to automatic speech recognition in a noisy environment yields good results. Keywords| Blind separation of convolutive mixtures, speech recognition in noisy environment.

In this study, the authors propose a new tracking algorithm for multistatic sonar systems, where the measurements collected by different sensors are sent to a fusion centre. The proposed algorithm relies on the main idea behind the trackbefore-detect paradigm, which consists of processing data from several consecutive pings, and estimates target positions by maximising the likelihood function of the available measurements. The authors assume that one manoeuvring target is present within the surveillance area. The preliminary performance assessment, carried out on simulated scenarios, shows that the proposed algorithm has acceptable performance also when the probability of detection per sensor is low (in the order of 0.3) and measurement errors are significant.

Activated reconnaissance systems based on target illumination are of high importance for surveillance tasks where targets are nonemitting. Multistatic configurations, where multiple illuminators and multiple receivers are located separately, are of particular interest. The fusion of measurements is a prerequisite for extracting and maintaining target tracks. The inherent ambiguity of the data makes the use of adequate algorithms, such as multiple hypothesis tracking, inevitable. For their design, the understanding of the residual clutter, the sensor resolution and the characteristic impact of the propagation medium is important. This leads to precise sensor models, which are able to determine the performance of the surveillance team. Incorporating these models in multihypothesis tracking leads to a situationally aware data fusion and tracking algorithm. Various implementations of this algorithm are evaluated with the help of simulated and measured data sets. Incorporating model knowledge leads to increased performance, but only if the model is in line with the physical reality: we need to find a compromise between refined and robust tracking models. Furthermore, to implement the model, which is inherently nonlinear for multistatic sonar, approximations have to be made. When engineering the multistatic tracking system, sensitivity studies help to tune model assumptions and approximations.

No abstract

No abstract