Moving vehicle identification using artificial neural networks

Abstract

Identification of moving vehicle type and its position using an acoustic sound signal is a critical task in many applications such as tactile sensory aid for differentially hearing ability abled, war field surveillance, border monitoring, traffic control systems. Moreover a wide range of research has been done for moving vehicle type identification using different signal processing techniques and algorithms. Differ from the previous researches this work intends to identify the moving vehicle type and its position from the subject. This research work is mainly divided into three sections. The first section deals with the state of the art review of various methods used to identify moving vehicles. The second section describes the experimental protocol design and feature extraction methods. Third section explains the classification of moving vehicles using Artificial Neural Network. A simple experimental protocol has been developed to record the moving vehicle sound signal using Sony recorder ICD-SX700. The recorded sound signals are preprocessed using the windowing technique. Feature extraction methods are based on two different approaches such as time domain and frequency domain. For time domain feature extraction, Autoregressive (AR) model and statistical features such as energy, kurtosis, standard deviation and skweness are used. For frequency domain, the signal is preprocessed using designed Butterworth filter with the band frequencies of 1/3 octave. The time domain signal is then transformed into frequency domain using Discrete Fourier Transform. The extracted features are then normalized and randomized to rearrange the values into definite range. Neural network models are used for the classification of moving vehicles. Network model such as Multilayer Perceptron (MLP), Radial Basis Function (RBF), Elman Neural Network (ENN) and Probabilistic Neural Network (PNN) are modelled. From the evaluation for the vehicle type and position classification, it can be inferred that the PNN model developed using the spectral band features has better classification accuracy while compared to other models. The performances of the networks are calculated and the results are compared.