Vehicle Classification Based on Seismic Signatures with Weighted Intrinsic Mode Functions

TL;DR

This paper introduces a weighted intrinsic mode function de-noising method combined with an improved MFCC and neural network for seismic vehicle classification, effectively reducing noise and improving recognition accuracy.

Contribution

It proposes a novel weighted EMD-based de-noising algorithm and integrates it with an enhanced MFCC and neural network for better seismic vehicle classification.

Findings

Weighted EMD improves signal clarity over traditional methods

Enhanced MFCC and neural network achieve higher classification accuracy

Method validated on DARPA's seismic vehicle data

Abstract

Seismic signal is used for vehicle classification widely. However, this task becomes difficult as a result of various noises. To solve the problem, this paper proposes a novel de-noising algorithm which evolves from a nonparametric adaptive tool named empirical mode decomposition (EMD). EMD can decompose signals into a set of zero-mean modes called intrinsic mode functions (IMFs) that can be used to denoise a signal. Unlike other EMD-based de-noising techniques, selecting the noise-free modes to denoise signals, this paper assigns appropriate weights to the modes. In addition, considering the similarities between speech recognition and seismic vehicle classification, an algorithm scheme, consisting of improved Mel frequency cepstral coefficient (MFCC) and artificial neural network, is applied to recognize seismic signals for vehicle targets. The data from DARPA's SensIt project, which contains various seismic signatures from two different vehicle types, is used to evaluate the method. Through experiments, results demonstrate the efficacy of proposed algorithm as compared to traditional MFCC.