Traffic target detection and identification method based on sound vibration time-frequency characteristics and cross attention fusion mechanism

Abstract

The invention provides a traffic target detection and identification method based on sound vibration time-frequency characteristics and a cross attention fusion mechanism. The method comprises the following steps: firstly, acquiring sound vibration signals of different traffic targets by adopting a sound vibration sensor, and carrying out NLMS (normalized least mean square) noise filtering pretreatment; secondly, variational mode decomposition (VMD) is carried out on the preprocessed acoustic vibration signals, and a scale spectrum segmentation method and a summation fuzzy entropy minimum value method are adopted to decompose the preprocessed acoustic vibration signals into a plurality of intrinsic mode functions (IMF); thirdly, extracting a Mel spectrogram from the sound signal IMF, extracting a wavelet transform time-frequency diagram from the vibration signal IMF, performing CNN convolution pooling on the result, and further extracting sound vibration signal features of different traffic targets through a transformer encoder; and finally, coding is carried out by using a cross attention mechanism, sound signal features and vibration signal features are fused into new features, and normalization and over-fitting prevention processing are carried out by using a Softmax function and a Dropout function. The method has the advantages of being low in algorithm complexity, high in real-time performance and low in cost, and meanwhile the traffic target detection problem under the scenes of extreme climate, weather, light and the like is solved.

Images