Research On Acoustic Vehicle Type Recognition By Recurrent Neural Networks

Vehicle type recognition, a quite basic research direction in intelligent transportation system,has been concerned by researchers at domestic and abroad. After 40 years of development in vehicle recognition, there still exists some problems such as low recognition rate, poor anti-interference and difficulty in data collection and processing, thus resulting in a gap in the real application of vehicle recognition. With the achievement of Recurrent Neural Networks(RNN) in the field of voice recognition, considering the differences of the interior noise in different vehicles, a RNN based vehicle type recognition is proposed in this thesis.In order to extracting the feature information of vehicle noise from long-time background information, we emphasis on RNN which belongs to feedback neural network.According to the current development of RNN, a bi-directional Long Short Term Memory(LSTM) network is introduced in this thesis as the model of vehicle noise recognition.The hidden layer of the recognition model employs LSTM units with input gate, output gate and reset gate. Each unit constitutes the forward propagating layer and the back propagating layer which are two discrete special constructional layers. In addition, these units completely connects with the input layer. This connection can solve long-time step bidirectional background information reference problem and gradient explosion/vanishing problem. In the final recognition stage, Connectionist Temporal Classification(CTC),overlying output of the model,is adopted to obtain the final output labels through the maximum probability decoding methods. The label reduction algorithm can eliminate repetitive and invalid labels, so corresponding issues between labels and data can be avoided in process of recognition, which also simplifies the data processing.In the respect of obtaining the test data, this thesis adopts the method of random sample collection for each sample. Firstly, a starting point is randomly chosen from the data stream; Secondly, the starting points are gotten according to the uniform distribution;Finaly, a slice with a same predetermined length is then extracted from raw data. In the subsequent data pre-processing, this thesis still chooses the method in the traditional voice recognition though the Mel Frequency Cepstrum Coefficient(MFCC) to create the final training and testing data. The final experiment indicates that bi-directional LSTM and CTC output model is robust after cross-validation testing. Compared with the Gaussian mixture model and the traditional neural network recognition model, the proposed model has the higher recognition rate. Additionally, under the influence of different degree of noise, it can fulfill the basic requirements of the vehicle type recognition task.