Road traffic monitoring is important for intelligenttransportation, and researchers have begun to focus on thedetection of traffic events using acoustic information. In thispaper, we apply model fusion to traffic acoustic event classi-fication. First, an improved, two-channel convolutional neuralnetwork (CNN) model is proposed as the weak classifier forconstructing the fusion model. The mel-cepstral feature andits first-order and second-order difference are selected as theinput features. Six different input features are constructed aftersignal preprocessing and segmentation. Second, after trainingsix different CNN models, the voting method and support vectormachine (SVM) stacking method are used to construct thefinal fusion model. Experimental results demonstrate that thedetection rate of traffic acoustic events reaches 95.1%, whichis higher than that of traditional traffic detection algorithms.
展开▼
