Acoustic Modality Based Hybrid Deep 1D CNN-BiLSTM Algorithm for Moving Vehicle Classification

Abstract

The main challenging goals in acoustic modality based moving vehicle recognition system is to accurately classify the moving vehicle with minimum misclassification rate. This article proposes an acoustic modality-based hybrid deep 1D convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) technique for moving vehicle classification under two-wheeler, low, medium, heavy weight category and noise. The proposed algorithm automatically extracts the high-level feature sequentially from experimentally generated vehicles signal and hold these features into the network for analyzing the time-varying characteristic for classification. Additionally, it is tested on the reference dataset SITEX02 for validation with 96% accuracy. The performance of 1D CNN-BiLSTM model has been compared with the conventional classifiers i.e., SVM, ANN, CNN and CNN-LSTM models. The experimental results show that CNN-BiLSTM has attained higher classification accuracy, 0.92 and minimum misclassification rate, 0.08 as compared to conventional classifiers. It has not only achieved satisfactory performance measures values and ROC-AUC characteristics but also obtained better generalization ability along with stability on the learning curve.