Research On Sound Recognition And Location Algorithm Based On Microphone Array

With the rapid development of urban economy in recent years, the traffic noise pollution has been becoming a crucial issue. Among the noise pollution, car horns is especially serious, and is drawing more and more attention. It is significant to supervise the illegal car horns with the sound source identification and localization methods, and put them into further practical application.The sound source localization (SSL) remains as a challenging task due to some factors, such as ambient noise impact and real-time calculation requirement. Existing methods can be roughly categorized into two groups. The first one is indirect method based on estimating the time differences of arrival (TDOA) of the signals received by spatially separated microphone pairs, and then calculate the source position according to the TDOA estimates and array geometry. The second one is the direct method. It steers the microphone array to various locations and then searches the peak of the output power, or analyzes the spatial-spectral correlation matrix derived from the signals, such as the steered response power (SRP) approach. This paper aims at combining the merits of these two groups of SSL algorithms, and proposes an efficient SRP-based method by using the strategy of stochastic region contraction (SRC) and simultaneously utilizing the TDOA estimates fully as well.In this paper, a novel system for car horn identification and localization which aims to helping the agency to monitor the illegal car. The main goal is to identify the existence of car horns from serious traffic interfering noise and then to get their accurate locations quickly.Firstly, in stage of identification, some short-term feature vectors, such as Mel frequency cepstrum coefficients (MFCC) and some other features in frequency domain are synthesized from the signals received by the microphone array in order to achieve good discrimination. Then, the back propagation neural network (BPNN) is applied to do the feature mapping for car horn recognition.Next, in the stage of localization, the steered response power (SRP) values and the time difference of arrival (TDOA) estimates are calculated respectively, for each independent microphone pair. Given the original searching region, we formulate the localization problem as a match-searching process in a spatial dictionary. The orthogonal matching pursuit algorithm (OMP) is implemented for match-searching during the process of region contracting in stochastic region contraction (SRC) strategy to obtain the location of car horn. When the searching region contract to a desired level or a predefined volume is achieved with several iterations, the location estimates of the target source can be calculated out.Both simulation and experiment results indicate that the proposed method can reduce the computations and achieve better time efficiency, with fewer iterations and faster contracting process, compared with the original SRC, which confirms the validity of the proposed method.