The invention relates to an abnormal public place sound extraction and recognition method for improving extreme-point symmetric mode decomposition (ESMD), abbreviated as D-ESMD. The method is characterized in that random T distribution sequence signals are added to abnormal public place sound, thereby reducing the influence of public place background noise on abnormal sound feature extraction; aiming at the problem of poor decomposition effect of original ESMD when abnormal sound is decomposed, a method for replacing extreme midpoint odd-even interpolation with symmetric midpoint interpolation is put forward to improve the decomposition efficiency and recognition rate of abnormal sound; aiming at the defects of original ESMD on effective decomposition mode selection, a permutation entropy algorithm is put forward to detect the complexity of the mode obtained by ESMD decomposition, so that an effective mode component of abnormal sound is adaptively obtained. By using the method, the feature of abnormal sound can be sufficiently described, good class recognition results can be obtained, and the feature of abnormal sound can be accurately extracted; and the method has good robustness on environmental background noise.