Vehicle acoustic recognition based on lightweight CNN
Vehicle type recognition technology is of great significance for traffic monitoring. Aiming at the problems that visual information of moving vehicles is susceptible to be disturbed by environment,a vehicle acoustic feature recognition model S-ShuffleNet based on ShuffleNet V2 is proposed,which contains the reduction of the network depth of ShuffleNet V2 and the improvement of the convolutional kernel size in the depthwise separable convolution,the log-Mel spectrogram features of the vehicle acoustic signal are extracted as the input of the model for vehicle recognition. Meanwhile,wind,rain,and thunderstorm noise are superimposed on the vehicle audio to verify the effect of different environmental noises. The experimental results demonstrate that the model has a low number of parameters and high training speed,compared with the basic network ShuffleNet V2,the recognition speed is improved by 2. 4％,and the recognition speed on the VS10 dataset can reach 97. 5％. Compared to other different classification networks,S-ShuffleNet also has good performance.
vehicle type recognitionacoustic featurelog-Mel spectrogram (LMS )lightweightconvolutional neural network(CNN)environmental noise
万方数据
维普
