A Multiple Sources Localization Method Based on TDOA Without Association Ambiguity for Near and Far Mixed Field Sources

A new method for multiple sources localization is proposed to eliminate association ambiguity for near and far mixed field sources. A spatial source localization model in the modified polar representation was constructed without the prior knowledge needed if the source is near-field or far-field. The localization model for multiple sources was deduced by using all the possible permutation of the TDOA sequences obtained from the original array by GCC-PHAT and the generalized trust region optimization processing method. In order to eliminate the phantom sources in the multiple sources localization, a set of calibration sub-arrays was constructed by switching the reference microphone in the array. The TDOA sequences of the estimated possible sources and all actual sources to the calibration sub-arrays were calculated separately. A reliability evaluation function was constructed based on the two sets of TDOA sequences, as well as a reliability evaluation between the sources identified by all the calibration sub-arrays. According to the principle of minimization of the reliability evaluation function, the real sources were screened out to solve the association ambiguity. Comparison analyses through simulation and experiment on real speech datasets were carried out under different localization scenarios. The results of simulation are consistent with the experimental results, which show that the proposed method effectively eliminates the phantom sources, and has higher positioning accuracy and robustness than the comparison methods, no matter sources are in the near-field or far-field.