2.3
Acoustic and Seismic Vehicle Recognition
There is a brief record of a study on automated moving vehicle localisation using acoustic and seismic signals, which was included in a report of the meeting held in the late 1970s [Brooks et al., 1977]. It was stated that they were interested in the velocity change of, and also the Doppler effect observed in a narrow-band spectra of acoustic and seismic signals. Although there are no journal publications found on this study, it indicated that the idea of combining acoustic and seismic signals processing for automated moving target detection had been considered.
2.3.1
Research on Military Vehicle Recognition
(a) [Altmann et al., 2002; Altmann, 2004]
Research on acoustic and seismic vehicle pattern recognition for security purposes was conducted by Altmann and his team in Germany. They collected a set of acous- tic and seismic data, collected from 10 different vehicles running at 7 various speeds on 4 separated lanes from 2 directions in a German armed force site, i.e. 560 data variations in total. They spent one month in 2000 to collect these and meteorological records as well as several recordings of background noise, including non-coloured noise. Their recognition algorithm utilised LVQ performed on the relative power within the first 15 harmonic lines observed in the power spectrum obtained by 2048 points FFT. The maximum recognition accuracy was reported to have reached 98% for four tracked vehicles and 83% for four wheeled vehicles.
It was reported that the classification of the heavy vehicles was relatively more successful; however for lighter vehicles, the algorithm required manual adjustment by the operator. It was also pointed out that the acoustic signal processing algo- rithm was developed with a focus more on the engine cycle rather than vehicle type. Moreover, in the more recent study, it was found that combining a beamforming algorithm to remove background noise by using microphone arrays can be useful.
2.3. ACOUSTIC AND SEISMIC VEHICLE RECOGNITION
However, beamforming was found to be useless for the seismic signal processing. Consequently, it was concluded that a further development of the algorithm is still necessary to improve its performance, especially for vehicle type recognition.
(b) [Li et al., 2002]
Li and his team published their experimental results of acoustic and seismic 2-class military vehicle type recognition, a classification study between tracked vehicles and wheeled vehicles, in IEEE Signal Processing Magazine. The PSD feature vec- tors were collected with FFT, from acoustic and seismic signals separately. The classification results by three algorithms, again on acoustic and seismic signals separately, were shown. The employed classifiers were: kNN, ML with GMM, and Support Vector Machines (SVM). It was understood that the SVM classifier achieved on average 94% for acoustic and 96% for seismic classification, and there- fore outperformed the other two.
(c) [Qu et al., 2003]
Qu and others [Qu et al., 2003] published a conference paper on their acoustic and seismic military vehicle classification study. Although the paper explained little about their algorithm and the experiment, it was understood that they used PSD combined with narrow band energy for feature extraction, and fuzzy logic rule based classifier and NN with Backpropagation learning algorithm. It was read that the average classification accuracy achieved during the experiment was approximately 85%.
(d) [Duarte and Hu, 2004]
Duarte and Hu reported their research on 4-class vehicle recognition using wireless distributed acoustic and seismic sensor network. The following were provided at each network node: a microphone, a geophone, an infrared sensor, together with a
2.3. ACOUSTIC AND SEISMIC VEHICLE RECOGNITION
processor that performed primarily recognition based on the local data, a transmitter and batteries. The data used in the research were collected at a USA military site over two weeks of November 2001 when they gathered raw data for later evaluation at each sensor node while also locally extracted feature vectors together with class labels that were manually designated by operators using their perception. During their experiment, local classification was performed only on the data segments, of which the presence of vehicles were detected by either the manual analyses or kNN algorithm; by the use of distance information and acoustic energy values. The local classification results of three classifiers; kNN, ML with normal distribution, and SVM were compared. The results showed similar levels of accuracy reached just under 70% by all three classifiers regarding acoustic data, and also slightly inferior performance on seismic classification; in which ML and SVM produced fairly better results than that of kNN.
(e) [Mazarakis and Avaritsiotis, 2006, 2007]
Mazarakis and Avaritsiotis attempted a two-class military vehicle classification with acoustic and seismic Wireless Sensor Networks (WSN). The chosen feature ex- traction algorithm was similar to Time Domain Signal Coding (TDSC) and Time- Encoded Signal Processing And Recognition (TESPAR), which was originally de- veloped by King and Phipps [King and Phipps, 1999] based on the Time-Encoded Speech (TES) research [King and Gosling, 1978]. TDSC and TES are explained later in Section 5.1.5. Results of two classifiers were compared; one was the commercially available Fast Artificial Neural Networks (FANN) that used multi- layer feedforward NN with Backpropagation algorithm, and another was called “Archetype” classifier that measured the distance between the average values of training sample of known class and a test sample regarding matrix representation of time domain signal shapes, developed for TDSC [Chesmore, 2001]. For analysis, they employed signals obtained from 9 runs each of two vehicles: a heavy wheeled
2.3. ACOUSTIC AND SEISMIC VEHICLE RECOGNITION
truck and a tracked vehicle. The non real-time classification experiment showed re- sults reaching around and above 80%, encouraging for time domain recognition studies, particularly combined with the proven simplicity and inexpensive com- putaion
(f) [Xiao et al., 2006]
Xiao and his team published two articles regarding their military vehicle recogni- tion study using acoustic and seismic signals. It was a comparative study between combinations of two feature extraction, two dimensionality reduction, and two clas- sification algorithms. For feature extraction, they chose STFT and a newly proposed method based on MFCC. Compared with the standard MFCC algorithm, theirs used rectangular filterbank instead of triangular, and there was no DCT included. The di- mensions of collected features were reduced by either Genetic Algorithm (GA) or PCA. Firstly, the two classifiers, kNN and SVM with RBF kernel for training, were processed on STFT with no dimensionality reduction algorithm, and it was found that SVM produced better accuracy than kNN.
For SVM and kNN, the effect of five different sizes of windowing function on classi- fication results was also analysed. Interestingly it was reported that the best window size was not the same for each classification algorithm when acoustic signals were used although it turned out to be the same for seismic recognition. This indeed indi- cated the importance of examining the best window size so as to improve the overall recognition performance. The rest of the experiment was only carried out with SVM algorithm using a fixed window size. It was reported that, on the whole, classifi- cation accuracy of acoustic recognition was better than that of seismic one because the acoustic classification achieved above 85% at maximum whereas the maximum accuracy for the seismic classification was around 70% up to 77%. Furthermore, they also mentioned that it would not be trivial to realise a fair comparison regarding results achieved so far in various vehicle recognition studies becasue of the multiple