Abstract: The traditional classification method based on Mel cepstrum coefficient and Gaussian mixture model is sensitive to interference noise in the classification process of field vehicles. To address the issue, an improved method based on dense convolution network structure (DenseNet) is proposed in this paper. First, the acoustic signal is converted to the spectrogram and then inputs to the improved DenseNet network structure for identification. The improved DenseNet network structure adds the function ‘center loss’ at the full connection layer to make the similar features more highly aggregated, so that the depth features of the acoustic signal can be extracted, which is beneficial to classification. The experimental results show that under the same sample set, the recognition rate of the improved DenseNet method can reach 97.70%, which outperforms the existing method.