Vehicle Class Recognition Based on Vehicle's Sound Signal

The focus of this paper is on classification of different vehicles using sound emanated from the vehicles. In this paper,quadratic discriminant analysis classifies audio signals of passing vehicles to bus, car, motor, and truck categories based on features such as short time energy, average zero cross rate, and pitch frequency of periodic segments of signals. Simulation results show that just by considering high energy feature vectors, better classification accuracy can be achieved due to the correspondence of low energy regions with noises of the background. To separate these elements, short time energy and average zero cross rate are used simultaneously.In our method,we have used a few features which are easy to be calculated in time domain and enable practical implementation of efficient classifier. Although, the computation complexity is low, the classification accuracy is comparable with other classification methods based on long feature vectors reported in literature for this problem.

Bruxism is a parafunctional orofacial behavior. For diagnosis, wearable devices that use sounds as biomarkers can be applied to provide the necessary information. Human beings emit various verbal and nonverbal sounds, making it challenging to identify bruxism-induced sounds. We wanted to investigate whether the acoustic emissions of different oral behaviors have distinctive characteristics and if the placement of the transducer has an impact on recording the sound signals. Sounds from five oral behaviors were investigated: jaw clenching, teeth grinding, reading, eating, and drinking. Eight transducers were used; six were attached to the temporal, frontal, and zygomatic bones with the aid of medical tape, and two were integrated into two commercial earphones. The data from 15 participants were analyzed using time-domain energy, spectral flux, and zero crossing rate (ZCR). In summary, all oral behaviors showed distinct characteristic features except jaw clenching, though there was a peak in the recording, possibly due to tooth tapping, before its expected onset. For teeth grinding, the transducer placement did not have a significant impact (p > 0.05) based on energy, spectral flux, and ZCR. For jaw clenching, the transducer placement had an impact with regard to spectral flux (p < 0.01). For reading and eating, the transducer placement had a significant impact with regard to energy (p < 0.05 for reading, p < 0.01 for eating), spectral flux (p < 0.001 for reading, p < 0.01 for eating), and ZCR (p < 0.001 for both reading and eating). For drinking, the transducer placement only had a significant impact with regard to ZCR (p < 0.01). We were able to record the sounds of various oral behaviors from different locations on the head. However, the ears were an advantageous location to place the transducer, since they could compensate for various head movements and ear devices are socially tolerable.

At a time of compression, two sound waveform segments each having single pitch length are cut-out from an input sound waveform at a time point represented by a current pointer and at a time point advanced from the time point by a single pitch period, respectively, and then, by adding the two sound waveform segments to each other after being multiplied by window functions, a single sound waveformn segment being compressed is produced. Next, the pointer is moved on the input sound waveform according to a compression rate, and then, a similar operation is repeated to produce a sound signal being compressed. At a time of expansion, two sound waveform segments each having double pitch length are cut-out from the sound waveform thus compressed at a time point represented by the current pointer and at a time point delayed from the time point by the single pitch period, respectively, and then, by adding the two sound wave segments to each other after being multiplied by window functions, a single synthesized sound wave form segment is obtained. Next, the pointer is moved on the sound waveform being compressed according to an expansion rate, and then, by repeating a similar operation, a sound signal being expanded is obtained.

Speech and Music Classification and Separation: A Review

To address the limitations of bridge expansion joint damage detection, this paper proposes a novel damage detection method that combines sound signal processing and deep learning. First, two modal bridge expansion joint (MBEJ) specimens were designed and fabricated. The support bar of the MBEJ was subjected to four damage states: undamaged, one-third damaged, two-thirds damaged, and completely damaged. Subsequently, a recorder was installed on the moving load to capture the sound signals generated as the load excited the support bar. Then, four features of the sound signal were extracted for analysis, including the short-time energy, short-time zero crossing rate, mel frequency cepstral coefficient (MFCC), and continuous Wavelet Transform (CWT). Finally, the features were fed into pre-trained and fine-tuned deep convolutional neural networks (DCNNs) for analysis, and a comparative study of the performance of different DCNN models was conducted. The results show that the features of the sound signal detect changes in the state of the support bar. The proposed method accurately classifies the state of the support bar using a single recorder. Optimal classification performance is achieved by the DCNN-based transfer learning model with MFCC and CWT features. For instance, the GoogLeNet model reached 100% and 99.8% accuracy with MFCC and CWT features, respectively.

The purpose of this study is to provide a systematic approach to address the industrial challenge of recognizing impeccable and defective gears by analyzing the spectrum of sound waves produced by the gears when in operation if attached to a testing circuit during sorting. This work classifies the gears into two classes, impeccable and defective classes. The spectra from several samples from both impeccable and defective gears were analyzed and five audio features were extracted from their spectra namely short-time energy, zero-crossing rate, Spectral entropy, pitch, and block energy entropy. It was found that there is a significant difference between the two classes. In training the algorithm, 5D features vectors from 20 feature vectors from impeccable gears and another 5D features vector from 20 defective gears as training samples to determine the discriminating point. In testing the algorithm, 20 samples were extracted randomly from the impeccable and defective gears but whose status was clearly known by visual inspection. The results of gear status given by the algorithm were compared to that of visual inspection. The samples were classified by using the Support Vector Machine (SVM) learning classification approach. A promising efficiency of 95% was obtained.

This paper describes a real-time system for recognizing voice commands for resource-constrained embedded devices, specifically a PIC microcontroller. While most existing speech ordering support solutions rely on high-performance processing platforms or cloud computation, the system described here performs fully embedded low-power processing locally on the device. Sound is captured through a low-cost MEMS microphone, segmented into short audio frames, and time domain features are extracted (i.e., Zero-Crossing Rate (ZCR) and Short-Time Energy (STE)). These features were chosen for low power and computational efficiency and the ability to be processed in real time on a microcontroller. For the purposes of this experimental system, a small vocabulary of four command words (i.e., “ON”, “OFF”, “LEFT”, and “RIGHT”) were used to simulate real sound-ordering interfaces. The main contribution is demonstrated in the clever combination of low-complex, lightweight signal-processing techniques with embedded neural network inference, completing a classification cycle in real time (under 50 ms). It was demonstrated that the classification accuracy was over 90% using confusion matrices and timing analysis of the classifier’s performance across vocabularies with varying levels of complexity. This method is very applicable to IoT and portable embedded applications, offering a low-latency classification alternative to more complex and resource intensive classification architectures.

Dimensions of vehicle sounds perception

In this work, we have presented a novel simple scheme for classifying audio speech signals into male speech and female speech. In the context of content-based multimedia indexing gender identification based on speech signal is an important task. Some popular salient low level time-domain acoustic features which are very closely related to the physical properties of source audio signal like zero crossing rate (ZCR), short time energy (STE) along with spectral flux, a low level frequency domain feature, are used for this discrimination. RANSAC and Neural-Net has been used as classifier. The experimental result exhibits the efficiency of the proposed scheme.

The train plug door is the only way for passengers to get on and off. The reliability of the doors has a direct impact on passengers' safety and operational efficiency. In order to address the shortcomings of the post-analysis and poor real-time of current fault diagnosis methods for train plug doors, a fault diagnosis method based on sound recognition is proposed. To process the non-stationary sound signals, the empirical mode decomposition (EMD) method is applied to sound signal samples of train plug doors, and a series of intrinsic mode functions (IMFs) are obtained. Then, wavelet packet decomposition is utilized on each IMF to acquire more detailed information. And wavelet packet energy entropy features are obtained. The Fisher discrimination criterion is used to carry out a mathematical analysis to select the most significant features as discrimination features. Finally, multi-class support vector machine (multi-class SVM) is utilized to carry out classification and validation. And the prediction accuracy of the 67 test samples reaches 95.52%, which indicates the proposed fault diagnosis method for train plug doors is feasible. The proposed method also provides the possibility of automatic faults recognition.

Investigation of acoustic and visual features for acoustic scene classification

The article presents a method for speech/pause segmentation based on energy analysis of speech signal fragments using the Teager Energy Operator and Short-Time Energy. The nature of segmentation is a linear division of a speech signal into fragments, calculation of energy characteristic of fragments using the Teager Energy Operator, calculation of Short-Time Energy values for fragments of energy characteristic, and determining `speech/pause' state of fragments based on the calculated threshold values of Short-Time Energy. The research on the proposed method was carried out, and the effectiveness of speech/pause segmentation was assessed in comparison with the method based on energy analysis using Short-Time Energy. In accordance with the obtained research results, it was concluded that due to the good susceptibility of the Teager Energy Operator to changes in signal amplitude and frequency, the proposed method provides the increased efficiency of speech/pause segmentation by 5.26 % and 5.51 % for the first and second kind errors, respectively. The proposed speech/pause segmentation method can be used for signal processing in real-time speech applications.

Vehicle classification plays an important part in Intelligent Transport System (ITS). However, the existing vehicle classification methods are not very robust to various changes such as lighting, weathers, noises, and the classification accuracy has been requiring to be improved. Sparse Representation-based Classifier (SRC) is not sensitive to the shortage and damage of data, the feature selection methods, is robust to lighting changes of the images, and can achieve excellent performance in multi-class classification problems. Therefore, this paper presents a new method based on spatial location information and SRC. Firstly, taken the specific characteristics of vehicles into consideration, the features of HOG (Histogram of Oriented Gradient) and HU moments are extracted to characterize the property of vehicles. In addition, the spatial location information is added to HU moment features in this paper to improve its ability to distinguish and describe vehicles. Then, vehicles are classified into six classes (large bus, car, motorcycle, minibus, truck and van) by SRC. The performance evaluation is carried out on the dataset which consists of 13300 vehicle images extracted from the real highway surveillance videos. The experimental results show that, using the proposed method, the time cost of classification can be speeded up by about 1.57 times and the average classification precision can achieve 96.53%, which is drastically improved by more than 2.7% compared to other existing methods.

Aimed at applying the sound recognition technology on the vehicle categories identification in border surveillance system, a new method of feature extraction is proposed based on the Empirical Mode Decomposition (EMD). The sound signals produced by different vehicles are important information source in the recognition system. These sound signals are decomposed by EMD, and a series of Intrinsic Mode Function (IMF) are obtained. Then the IMFs are analyzed by Fast Fourier Transform (FFT). According to the obtained amplitude spectrum of each IMF, the corresponding energy are calculated and normalized, which are regarded as feature vectors. After that, neural network is designed to recognize different types of vehicle in border surveillance. The simulation results indicate that the identification ratio grows up to 89.7% using this feature extraction method, which improves 11.7% and 1.8% by comparing with the methods based on Short Time Fourier Transform (STFT) and Wavelet Transform (WT), respectively.

This paper utilizes a rigid spherical microphone array to reduce wind noise. In the experiments conducted, a loudspeaker is used to reproduce the desired sound signal and an axial fan is employed to generate wind noise in an anechoic chamber. The sound signal and wind noise are measured separately with the spherical microphone array and analyzed in the spherical harmonic domain. The wind noise is found to be irregularly distributed in the spherical harmonic domain, distinct from the sound signal which is concentrated in the first few spherical harmonic modes. This difference is utilized to reduce wind noise without degrading the desired sound pressure level (SPL) by use of a low pass filter method in the spherical harmonic domain. Experimental results with both single-tonal and multi-tonal sound signals demonstrate that the proposed method can reduce wind noise by more than 10 dB in the frequency range below 500 Hz. The SPL of the desired sound signal can be extracted from wind noise with an error within 1.0 dB, even when the sound level is 8 dB lower than wind noise.