GB2434876A

GB2434876A - Frequency and time audio signal discriminator

Info

Publication number: GB2434876A
Application number: GB0602032A
Authority: GB
Inventors: Julian Whiffen
Original assignee: Thales Holdings UK PLC
Current assignee: Thales Holdings UK PLC
Priority date: 2006-02-01
Filing date: 2006-02-01
Publication date: 2007-08-08
Anticipated expiration: 2026-02-01
Also published as: GB2434876B; GB0602032D0

Abstract

A discriminator (202) is disclosed for extracting features of an audio signal (210), comprising frequency analyser means (204) for processing the audio signal (210) to generate a plurality of frequency output signals, each frequency output signal being associated with a respective frequency band of the audio signal and representing a signal strength of said frequency band; and modulation analyser means (206) for processing the frequency output signals to generate a plurality of modulation output signals, each modulation output signal representing the variation over time of a corresponding frequency output signal. An audio signal classifier, intrusion detection system, container, carrier medium and method are also disclosed. The audio signal discriminator has particular application in an intrusion detection system suitable for use in shipping containers.

Description

AUDIO SIGNAL DISCRIMINATOR

The present invention relates to a discriminator, an audio signal classifier, an intrusion detection system, a container, a method of extracting features of an audio signal, and a carrier medium. The invention finds particular application in the field of audio source classification and the protection of secure volumes such as shipping containers.

It is often desired to protect a secure volume from intrusion or unauthorised access. This can be done, for example, by use of infra-red sensors which monitor thermal radiation in the volume and which raise an alarm if the thermal radiation changes in a predefined way (such as an increase in thermal radiation caused by an intruder entering the field of view of the sensor, for example). Another way of protecting a volume is by attaching contact sensors to entry and exit points of the volume, for example, and raising an alarm if the sensors are tripped' while the system is active. Other detection systems, such as pressure sensors, optical sensors, and the like, may also be used.

For certain applications such as the protection of cargo shipping containers, it is not practical or cost-effective to use such systems, and security is normally provided instead by essentially passive means (by providing strong metal walls and a door, and physically limiting access to the container using locks and the like).

Security for shipping containers and the like can nevertheless be a problem because intruders can break into such containers by drilling or sawing through the walls to make a hole, and breaking the locks of the containers (amongst other methods). These attacks can be relatively varied and unpredictable, and may not necessarily be prevented by additional internal security measures such the thermal sensors, contact sensors or optical sensors mentioned above, owing to problems with field of view (since the container is usually full of goods, blocking the view of any sensor) and because the attacks do not necessarily use a recognised entry/exit point, for example. Additionally, shipping containers (for example) are subjected to a wide variety of environmental stresses (vibration, movement and temperature variations) which could trigger false alarms in security systems such as those mentioned above. Attention has therefore tended to focus on providing external security measures, such as surveillance of the area in which a shipping container is stored (but with limited success).

During consideration of the problems posed by the effective protection of secure volumes such as shipping containers, it was noted that many typical intrusion attempts involve the production of noise within the volume. It was furthermore found that some noises (such as the sound of breaking glass) can in some cases produce a sound with a relatively clear and distinctive frequency spectrum.

However, it was also found that systems which monitored overall sound levels or which analysed the frequency spectrum of noises within the volume to be protected were generally not effective because of potentially high levels of background noise (for example caused by knocking or moving a container) and because noises such as drilling and sawing were not found to have a distinctive frequency spectrum. Furthermore, attacks such as drilling and sawing were found to produce different noises depending on the speed of the drilling or sawing, the type and pitch of drill bit or saw blade, and so on. It was found that because of factors such as these, identifying an audio source on the basis of an audio signal received within a secure volume was relatively difficult.

It is thus an object of the present invention to provide a discriminator for detecting diverse audio sources, such as intrusion attempts within a shipping container.

Accordingly, in a first aspect of the invention, there is provided a discriminator for extracting features of an audio signal, comprising: frequency analyser means (such as a processor and optional associated memory) for processing the audio signal to generate a plurality of frequency output signals, each frequency output signal being associated with a respective frequency band of the audio signal and representing a signal strength of said frequency band; and modulation analyser means (such as the same or a further processor and optional associated memory) for processing the frequency output signals to generate a plurality of modulation output signals, each modulation output signal representing the variation over time of a corresponding frequency output signal.

It will be appreciated that each of the audio signal, frequency output signals and modulation output signals may be digital (such as a digitised stream of sample data) or analogue (continuously varying values), as described in more detail below. The frequency analyser means and modulation analyser means may be implemented in a single component or device, or could be provided independently, for example, and may also include intervening processing stages. 12 frequency output signals may be provided, but more or less (even as few as 1) may be used.

By generating modulation output signals that represent the variation over time of the signal strength of various frequency bands of an input audio signal, it was found that sounds such as sawing and drilling could be identified on the basis of these modulation output signals, despite the possible presence of background noise and despite such sounds generally lacking a distinctive frequency spectrum.

The discriminator may further comprise output means (such as an input/output device) for outputting the modulation output signals. The discriminator may also or alternatively output only the frequency output signals, but it was found that classification of sounds could be carried out only on the basis of the modulation output signals, thus providing a relatively simple set of discriminator outputs.

The frequency analyser means may be adapted to generate each frequency output signal in dependence on the signal power within the corresponding frequency band. The power may be computed by taking the square of the magnitude of input audio samples (in either the time or the frequency domain, for example). Analysing the power, rather than magnitude, of the signals can reduce the impact of background noise. Alternatively only the signal amplitude may be used, in order to simplify the processing and to reduce computational load. (The term signal strength' as used herein preferably comprises both signal power and signal magnitude.) The frequency analyser means may also be adapted to generate each frequency output signal in dependence on an average of the signal strength within each frequency band.

This may be by use of a low-pass filter or similar, using a 3db cut-off frequency that is typically not greater than the bandwidth of the corresponding frequency band. This averaging feature can produce a more flat and uniform frequency output signal, leading to more effective discrimination. The averaging step is preferably undertaken after computing the spectral power of the audio signal (see above).

The frequency analyser means may be adapted to generate each frequency output signal in dependence on a substantially logarithmic function (such as a natural or other logarithm, a Sigmoid, A-law or u-law function) of the signal strength within the corresponding frequency band. Any essentially non-linear function may alternatively be used. This can provide a larger dynamic range and simulates the sensitivity of human hearing. This step is preferably carried out after the averaging step (see above).

The frequency analyser means may also be adapted to analyse frequency bands that have centre frequencies forming a substantially exponential series. That is, the centre frequency f of a frequency band n may be of the form such as = A. eB l, whereby a choice of constants such as A=64 Hz and B=O.44 will give rise to centre frequencies of approximately 100 Hz, 160 Hz, 240 Hz, 380 Hz, 580 Hz, 900 Hz, 1.4 kHz, 2. 2 kHz, 3.4 kHz, 5.3 kHz, 8.2 kHz and 12.6 kHz, for example, for a frequency analyser having twelve different frequency bands. Alternatively, a linear spacing of frequency bands, where the centre frequencies of the bands have the formf = C.n (such as 1 kHz, 2 kHz, 3 kHz, and so on) may be used for simplicity. The benefit of an exponential/logarithmic frequency scale allows coverage of a wide range of frequencies (for example from 100 Hz to 12.6 kHz) yet provides a relatively fine resolution at low frequencies. This frequency scale also corresponds approximately to the sensitivity of the human ear.

The frequency analyser means further comprise band-pass filtering means (such as a dedicated filter or filter bank, or a processor and optional associated memory) for selecting each of the plurality of frequency bands for processing essentially independently of the other frequency bands. This can simplify the processing and filtering operations. In more detail, the band-pass filters may be configured such that the 3db cut-off frequencies of adjacent band-pass filters are at essentially equal frequencies.

The filter designs may be optimised to achieve a desired trade-off of speed, storage capacity (of filter coefficients) and accuracy.

Each modulation output signal may represent a signal strength within a frequency band of the corresponding frequency output signal. The modulation analyser means may, moreover, be adapted to generate multiple modulation output signals per frequency output signal, each of said multiple modulation output signals being associated with one of a corresponding plurality of frequency bands of the frequency output signal. Thus, for a particular frequency band there may be provided a number of modulation output signals relating to different frequencies of modulation (of power and so on) within the frequency band. Eight different modulation outputs may be used per frequency band (frequency output signal), but more or less (even as few as 1) may be used.

As with the frequency analyser, the modulation analyser means may be adapted to generate each modulation output signal in dependence on the signal power within the corresponding frequency band of the corresponding frequency output signal. Again, the signal amplitude or other measure may be taken. Again as with the frequency analyser, the modulation analyser means may be adapted to generate each modulation output signal in dependence on an average of the signal strength within the corresponding frequency band of the corresponding frequency output signal, and the modulation analyser means may be adapted to generate each modulation output signal in dependence on a substantially logarithmic function of the signal strength within the corresponding frequency band of the corresponding frequency output signal.

The modulation analyser means may also be adapted to analyse frequency bands that have centre frequencies forming a substantially exponential series. The centre frequencies may be chosen such that the highest centre frequency used by the modulation analyser is comparable to twice the highest frequency of noise likely to be generated by devices used for intrusion attacks (100 Hz, for example, for devices producing noise not exceeding 50 Hz). The lowest centre frequency may be 0.5, 1, 2, 3, or 10 Hz for example (or higher or lower). This feature again allows coverage of wide range of frequencies whilst retaining a relatively fine resolution at low frequencies The modulation analyser means may further comprise band-pass filtering means for selecting each frequency band (of the relevant frequency output signal) for processing essentially independently of any other frequency bands.

In another aspect of the invention there is provided an audio signal classifier for classifying an audio signal, comprising: input means (such as an audio connector andlor microphone) for inputting the audio signal; a discriminator as aforesaid for extracting features of the audio signal; classifier means, adapted to process signals output by the discriminator to identify an audio signal source likely to be associated with the audio signal; and output means for outputting an identification signal representative of the identified audio signal source. The classifier means may be a computer-implemented neural network, and may be trained with a number of test attack' and non-attack' sounds. Alternatively the classifier means may be adapted to compute eigenvectors and eigenvalues of the discriminator outputs, andlor perform principal components analysis and the like, for example. This can allow an audio signal source to be identified in response to inputting a single audio signal into the device.

The classifier means may further comprise means for accessing (such as an I/O device optionally including storage means such as a hard disk) audio signature data associated with a plurality of possible audio sources, and means for processing the audio signature data in conjunction with the signals output by the discriminator to identify the audio source most likely to cause the signals output by the discriminator. This can improve the detection rate of the device, as it can focus on a library of known sounds. Alternatively or additionally the audio signature data may be altered or added to during use, for example to provide a training' function for the device to allow it to recognise new audio sources and the like.

In a further aspect of the invention there is provided an intrusion detection system for use in a protected space, comprising: an audio signal classifier as aforesaid; and intrusion detection means for processing the identification signal output by the audio signal classifier, and for outputting an alarm signal if the identified audio source relates to an intrusion into the protected space. This can provide a relatively low-cost intruder alarm system suitable for use in shipping containers and the like.

Accordingly, in another aspect of the invention there is provided a container including an intrusion detection system as aforesaid.

In a further aspect of the invention there is provide a method of extracting features of an audio signal, comprising: processing the audio signal to generate a plurality of frequency output signals, each frequency output signal being associated with a respective frequency band of the audio signal and representing a signal strength of said frequency band; and processing the frequency output signals to generate a plurality of modulation output signals, each modulation output signal representing the variation over time of a corresponding frequency output signal. Further method steps may be provided in accordance with the apparatus equivalents.

The present invention can be implemented in any convenient form, for example using dedicated hardware, or a mixture of dedicated hardware and software. The present invention is particularly suited to implementation as computer software implemented by a Digital Signal Processor, microcontroller system, or a workstation or laptop computer.

The invention may further comprise a network (for example for transmitting an alarm signal), which can include any local area network or even wide area, conventional terrestrial or wireless communications network. The systems may comprise any suitably programmable apparatuses such as a general purpose computer, personal digital assistant, mobile telephone (such as a WAP or 3G-compliant phone) and so on. Aspects of the present invention encompass computer software implementable on a programmable device. The computer software can be provided to the programmable device using any conventional carrier medium. The carrier medium can comprise a transient carrier medium such as an electrical, optical, microwave, acoustic or radio frequency signal carrying the computer code. An example of such a transient medium is a TCP/IP signal carrying computer code over an IP network, such as the Internet.

The carrier medium can also comprise a storage medium for storing processor readable code such as a floppy disk, hard disk, CD ROM, magnetic tape device or solid state memory device.

Although each aspect and various features of the present invention have been defined hereinabove independently, it will be appreciated that, where appropriate, each aspect can be used in any combination with any other aspect(s) or features of the invention.

Embodiments of the present invention will now be described with reference to the accompanying drawings, in which: Figure 1 is a schematic of a shipping container including an intruder detection device in accordance with an embodiment of the present invention; Figure 2 is a schematic diagram of discriminator and classifier components of the intruder detection device of Figure 1; Figure 3 is a schematic diagram of frequency analyser and modulation analyser components of the discriminator component shown in Figure 2; Figure 4 is a schematic diagram of the classifier component of Figure 2 in more detail; Figure 5 is an overview of a hardware implementation of the system shown in Figure 2 to 4; Figure 6 is a plot of filter design characteristics of band-pass filters used in the frequency analyser shown in Figure 3; and Figure 7 is a plot of filter design characteristics of band-pass filters in the modulation analyser shown in Figure 3.

An intrusion detection system, incorporating a discriminator for extracting features of audio signals, will now be described.

Figure 1 shows schematically a shipping container 100 having walls 102 and a lockable door 104 which form a secure volume for storing cargo and other materials to be transported or stored in the container. Through the cut-away section indicated with curved lines can be seen an intrusion detector system 106 and a microphone 108 for capturing acoustic emissions within the container 100. The walls 102 and door 104 are metal and can be locked shut andlor sealed, preventing casual unauthorised access to the container.

Access to the contents of the container is nevertheless possible by breaking the locks to the container or by sawing andlor drilling through the walls 102 or door 104 to form a hole in the walls or door through which the contents of the container can be removed.

Sawing and drilling into the sides or rear of a shipping container can be relatively difficult to detect by external surveillance of the area in which the container is stored (typically together with many other containers).

The intrusion detector 106 detects intrusions into the container from inside the container. The detector 106 continuously monitors audio emissions within the container via the microphone 108, and raises an alarm if it detects an acoustic emission associated with an intrusion attempt, such as the sound of a saw blade or a drill bit being driven through metal. The alarm can be acoustic (such as a loud alarm sound), electronic (a wireless or other transmission to cause an external alarm system to be activated) or otherwise.

The intrusion detection system will now be described in more detail with respect to Figure 2.

Figure 2 shows schematically the components of the intrusion detection system described above. The detection system 200 includes a discriminator 202, including a frequency analyser component 204 and a modulation analyser component 206, and a classifier 208. The classifier is connected in turn to an alarm system (not shown) for raising the alarm if an intrusion is detected. The output of the microphone is fed into the frequency analyser 204 via audio input 210.

As is described in more detail below, the frequency analyser performs a spectral analysis of the audio signal 210, and produces a set of outputs (referred to herein as frequency output signals') corresponding to the signal power within each of 12 separate frequency bands. The output of the frequency analyser is in some respects comparable to taking a Fourier transform of the audio input signal 210, but with some differences, as explained below. The modulation analyser 206 then produces 8 outputs (referred to herein as modulation output signals') for each frequency output signal.

Each of the modulation output signals for a particular frequency output signal provides information on the degree to which the frequency output signal is varying within a particular frequency band. In total, the modulation analyser provides 96 outputs (8 x 12) providing an indication how each of the 12 different frequency bands of the input audio signal 210 are varying over time. (It was found that many typical intrusion noises such as sawing, drilling, and so on could be more easily identified by analysing how frequencies varied over time than by analysing frequencies in a snapshot' of the audio signal.) The 96 modulation output signals from the modulation analyser 206 are then fed into the classifier 208 (see below) for classification into likely audio sources.

The human ear is responsive to sounds having a frequency from approximately 100 Hz to 20 kHz, and the discriminator analyses sounds approximately within this region. In order to allow the number of frequency channels to be minimised whilst maintaining a relatively high resolution at low frequencies (where much of the information content of noises are present), the audio input signal 210 is split into frequency bands having an exponential/logarithmic relationship. In more detail, the frequency bands were chosen such that the centre frequency f of a frequency band n is of the form f, = A eB and constants of approximately A =64 Hz and B=O. 44 were chosen, giving rise to centre frequencies of approximately 100 Hz, 160 Hz, 240 Hz, 380 Hz, 580 Hz, 900 Hz, 1.4 kHz, 2. 2 kHz, 3.4 kHz, 5.3 kHz, 8.2 kHz and 12.6 kHz (as opposed to a linear scheme in which the audio frequency is divided linearly into 12 frequency bins' of width 1

kI-Iz, for example).

The modulation analyser also uses a substantially logarithmically/exponentially-spaced frequency distribution, dividing each incoming frequency output signal into 8 frequency bands having centre frequencies of 2 Hz, 3.5 Hz and so on up to 100 Hz.

The discriminator will now be described in more detail with respect to Figure 3.

Figure 3 shows schematically the components of the discriminator. The discriminator components 300 are divided into the frequency analyser 304 components and modulation analyser 306 components. Three of the 12 frequency analyser filter banks are shown. The top filter bank analyses the first frequency band, centred on 100 Hz. In the first (uppermost) filter bank, the incoming audio input 302 is first band-pass filtered with a centre frequency of 100 Hz (subsequent banks use band- pass filters with centre frequencies of 160 Hz, 240 Hz, 380 Hz, and so on). The power of the filtered signal is then computed, by squaring the output of the band-pass filter. The signal is then low-pass filtered (with a 3db cut-off point of 100 Hz) to provide an averaging effect, and a logarithm is then taken of the averaged power, to provide the frequency output signal 308. The same filtering is carried out for the remaining 11 frequency analyser filter banks (except for the different band-pass filter characteristics), producing the other frequency analyser outputs 310, and so on.

Each frequency output signal 308, 310 is then fed into 8 further filter banks in the modulation analyser 306. 3 of the 8 filter banks for the first frequency analyser output 308 are shown. The 8 filter banks are duplicated for each of the remaining 11 frequency analyser outputs 310, and so on, making a total of 96 filter banks in the modulation analyser 306, as noted above.

The first modulation analyser 306 filter bank includes a band-pass filter with a centre frequency of 2 Hz, to detect modulations (in the 100 Hz region of the audio signal) having a period of approximately half a second. Subsequent banks of the 8 filter banks for any given frequency analyser output use exponentially increasing centre frequency values, up to 100 Hz (detecting very fast modulations having a period of approximately 0.01 seconds). As in the frequency analyser, the power is then taken, the signal is averaged (using a low-pass filter with 3db cut-off point of 0.5 Hz), and the logarithm of the signal is taken.

The use of logarithms effectively improves the dynamic range of the system, and improves the performance of the system with regard to multi-path transmissions of sound.

An explanation of the operation of the system will briefly be given. In this example, intruders A, B and C attempt to break into a container using a hacksaw operated at approximately 2 Hz, another hacksaw operated at approximately 3.5 Hz, and a drill with a dominant frequency of approximately 12.6 kHz (modulating slightly over time) respectively.

Assuming that the hacksaw generates a wide range of frequencies (including frequencies at approximately 100 Hz), both intruders A and B may cause a relatively strong frequency output signal for 100 Hz (output 308). The noise associated with intruder A will then cause a relatively strong output at the output 312 (since the 100 Hz frequency is modulating in time with his sawing) but a weaker output 314, whereas the noise associated with intruder B will cause a relatively strong output at the output 314 and a weaker output at output 312 (since his sawing is at a different frequency).

However, overall the noises associated with both intruders will create a relatively distinctive pattern of relatively strong outputs at relatively low frequency modulations for all of the frequency bands of the audio signal.

By contrast, the noise made by intruder C may cause little effect at any modulation output signals relating to low frequencies of the audio signal, but may cause a relatively strong and characteristic response at the low frequency modulation outputs relating to the 12.6 kHz frequency band. Thus it will be appreciated that the system can not only detect but discriminate between different types of intrusion attempts that do not produce a distinctive frequency spectrum (when viewing a snapshot' of the noise).

The classifier component will now be described in more detail with reference to Figure 4.

Figure 4 is a schematic of the classifier system 400. The classifier 402 includes a classification module 404 and a decision module 406. The classification module 404 receives the outputs 408 (the 96 different modulation output signals) of the discriminator module mentioned above, and processes the outputs 408 to generate classification signals 410 representative of the likelihood that a number of different audio sources are present in the detection volume. The decision module 406 processes the classification signals 410 to determine whether a particular audio source relating to an intrusion attempt is present and, if so, transmits an alarm signal 412 to raise the alarm.

The classification module 404 comprises a neural network which has been trained to recognise a number of attack' and non-attack' noises. The number of layers and nodes in the neural network and the amount of training of the network can be varied depending on the available computing power, the desired accuracy and the number of different noises desired to be distinguished between. A neural network distinguishing between 11 different noises was found to function adequately with no more than 11 output nodes (each output node corresponding to a particular one of the noises), for example, but other configurations are of course possible. The classification signals 410 output by the classification module 404 represent the strength of the output of each of the output nodes of the neural network. The decision module 406 then determines whether a particular alarm noise is present by comparing each of the classificationsignals 410 to a detection threshold. The threshold can be raised or lowered depending on the desired compromise between the detection sensitivity and false alarm probability.

Different thresholds may be applied to different classification signals (noise types) if desired.

In a variant of the preferred embodiment, a "signal-space" classifier is used. In this variant, the classification module is trained by processing a fixed sample set (1000) of the 96 discriminator outputs. A 96-dimensional volume is then created corresponding to the sample volume into which 80% (or other proportion) of the 1000 samples are located, and the volume is characterised by 96 mean values (defining the centre of the volume) and 96 standard deviation (SD) values defining the corresponding widths of the volume.

In use, the classification module forms a 96-length vector of samples based on the values of the 96 discriminator outputs, and tests to see whether the vector lies within any of the predefined volumes. If the vector does fall within such a volume, a match is identified corresponding to the acoustic source associated with the relevant volume.

In this variant, because many of the modulation output signals may be correlated, the classification can be improved using a principle component analysis method (computing eigenvectors and eigenvalues of the sample set). Alternatively or additionally, a neural network, initially trained using a variety of attack and non-attack noises, can be used to carry out audio source classification. Other types of classifier system and training methods may of course be used.

The structure of the intrusion detector system will now be discussed briefly with reference to Figure 5.

Figure 5 is an overview of the intrusion detector system 500. The system comprises a processor 502, an input/output device 504, a program data store 506 and a data store 508. The processor carries out the filtering operations shown in Figure 3 and also carries out the classification process mentioned above with reference to Figure 4, as well as carrying out ancillary tasks such as monitoring system health and transmitting alarm signals if necessary. The system may also include components for networking andlor generating an audible or visible alarm signal (not shown) for example.

The input/output device handles data input from the microphone, and data output with the alarm device (or external system, for example via a network interface card). The program data store 506 contains the computer program code necessary to carry out the various operations of the intrusion detector (filtering, classification, and so on), and the data store 508 contains data such as acoustic signature data (if applicable), filter coefficients, audio sample storage, and so on.

The intruder detector system may alternatively be implemented using an ASIC (or the like) with dedicated hardware filter banks, for example.

The filter characteristics of the band-pass filters used in the frequency analyser and modulation analyser will be described in more detail.

Figure 6 is a logarithmic graph showing (typical) desired filter characteristics for the 12 band-pass filters used in the frequency analyser. The centre frequencies range from 100 Hz to 12.6 kHz, and have a response such that the top 3db cut-off point of one filter cross the bottom 3db cut-off point of the next filter (and vice versa). The filters shown were designed using a bi-linear z-transform with the usual tan() distortion only applied to the upper and lower frequency limits to preserve the crossing points. A side-effect of the hi-linear z-transform is that the filters are shaped effectively to block (inaudible) frequencies above 20 kHz. The filter constraints (the frequency range, and the requirement for 3 db points of adjacent filters to touch) resulted in a Q factor of 2.1.

Figure 7 is a logarithmic graph showing (typical) desired filter characteristics for the 8 band-pass filters used in the modulation analyser. The absolute powers of the signals within the frequency analyser filters is therefore not relevant, giving some measure of tolerance to the path (meaning multi-path) by which the signal reaches the microphone.

The filter constraints resulted in filters with a Q factor of approximately 1.77.

Double pole, critically damped filters were used for the low-pass filters in the frequency analyser component and the modulation analyser component mentioned above.

The intrusion detector described above can be used to trigger an alarm directly.

However, the system may alternatively or additionally be used to activate additional intruder sensor systems, such as the infra-red systems mentioned above. Furthermore, the output of the intrusion detector system described above can be combined with the output of other systems and fed into a majority voting or similar system so that, for example, the output of the system described herein can be made relatively sensitive (easy to activate) but the number of false alarms can be kept to a minimum.

With regard to the modulation analyser design, it is possible to include a low-pass filter with a 3dB point coinciding with the low-edge of the lowest frequency band-pass filter in the input filter bank. It was chosen not to do this in the embodiment described above because the system would be sensitive to the route taken by the signal to the microphone.

It should also be appreciated that whilst a certain number of frequency analyser channels and modulation analyser channels have been described above, different numbers of channels (more or less) may of course be provided. The channels may also be differentiated as appropriate. For example, some discriminator outputs may be essentially null outputs (conveying little or no information) because of the interaction between the frequency analyser and modulation analyser filter characteristics; the 100 Hz modulation analyser output for the modulation analyser channel attached to the 100 Hz frequency analyser output will be such a null output, for example. The modulation analyser and frequency analyser components may therefore be altered to remove null outputs (for example by reducing the number of channels in the 100 Hz modulation analyser channel). However, the symmetry of the system described above can allow simpler construction of the discriminator.

While the discriminator and associated systems have generally been described above in relation to protecting the security of shipping containers and the like from attack by drilling, sawing and so on, it will be appreciated that the system may be applied to any volume in which it is desired to detect specific sounds. For example, the system may be installed in a public car park and used to detect sounds relating to unauthorised entry into, or vandalism of, parked vehicles. Alternatively, the system may be used to detect other types of sound, not necessarily for the purpose of security. For example, the system could be used to detect different types of failure modes in buildings or vehicles, or to identify different types of bird song or animal calls in an outdoor space.

The system may also be expanded to include more than one microphone, and may be used to complement, or be combined with, other intrusion detection systems mentioned above.

Further modifications lying within the spirit and scope of the present invention will be apparent to a skilled person in the art.

Claims

CLAIMS: 1. A discriminator for extracting features of an audio signal,

comprising: frequency analyser means for processing the audio signal to generate a plurality of frequency output signals, each frequency output signal being associated with a respective frequency band of the audio signal and representing a signal strength of said frequency band; and modulation analyser means for processing the frequency output signals to generate a plurality of modulation output signals, each modulation output signal representing the variation over time of a corresponding frequency output signal.

2. A discriminator according to Claim 1, further comprising output means for outputting the modulation output signals.

3. A discriminator according to Claim 1 or 2, wherein the frequency analyser means is adapted to generate each frequency output signal in dependence on the signal power within the corresponding frequency band.

4. A discriminator according to any preceding claim, wherein the frequency analyser means is adapted to generate each frequency output signal in dependence on an average of the signal strength within each frequency band.

5. A discriminator according to any preceding claim, wherein the frequency analyser means is adapted to generate each frequency output signal in dependence on a substantially logarithmic function of the signal strength within the corresponding frequency band.

6. A discriminator according to any preceding claim, wherein the frequency analyser means is adapted to analyse frequency bands that have centre frequencies forming a substantially exponential series.

7. A discriminator according to any preceding claim, wherein the frequency analyser means further comprises band-pass filtering means for selecting each of the plurality of frequency bands for processing essentially independently of the other frequency bands.

8. A discriminator according to any preceding claim, wherein each modulation output signal represents a signal strength within a frequency band of the corresponding frequency output signal.

9. A discriminator according to Claim 8, wherein the modulation analyser means is adapted to generate multiple modulation output signals per frequency output signal, each of said multiple modulation output signals being associated with one of a corresponding plurality of frequency bands of the frequency output signal.

10. A discriminator according to Claim 8 or 9, wherein the modulation analyser means is adapted to generate each modulation output signal in dependence on the signal power within the corresponding frequency band of the corresponding frequency output signal.

11. A discriminator according to any one of Claims 8 to 10, wherein the modulation analyser means is adapted to generate each modulation output signal in dependence on an average of the signal strength within the corresponding frequency band of the corresponding frequency output signal.

12. A discriminator according to any one of Claims 8 to 11, wherein the modulation analyser means is adapted to generate each modulation output signal in dependence on a substantially logarithmic function of the signal strength within the corresponding frequency band of the corresponding frequency output signal.

13. A discriminator according to any one of Claims 8 to 12, wherein the modulation analyser means is adapted to analyse frequency bands that have centre frequencies forming a substantially exponential series.

14. A discriminator according to any one of Claims 8 to 13, wherein the modulation analyser means further comprises band-pass filtering means for selecting each frequency band for processing essentially independently of any other frequency bands.

15. An audio signal classifier for classifying an audio signal, comprising: input means for inputting the audio signal; a discriminator for extracting features of the audio signal, as defined in any of Claims ito 14; classifier means, adapted to process signals output by the discriminator to identify an audio signal source likely to be associated with the audio signal; and output means for outputting an identification signal representative of the identified audio signal source.

16. An intrusion detection system for use in a protected space, comprising: an audio signal classifier as defined in Claim 15 or 16; and intrusion detection means for processing the identification signal output by the audio signal classifier, and for outputting an alarm signal if the identified audio source relates to an intrusion into the protected space.

17. A container including an intrusion detection system as defined in Claim 16.

18. A method of extracting features of an audio signal, comprising: processing the audio signal to generate a plurality of frequency output signals, each frequency output signal being associated with a respective frequency band of the audio signal and representing a signal strength of said frequency band; and processing the frequency output signals to generate a plurality of modulation output signals, each modulation output signal representing the variation over time of a corresponding frequency output signal.

19. A method according to Claim 18, further comprising outputting the modulation output signals.

20. A method according to Claim 18 or 19, wherein the step of processing the audio signal includes the step of generating each frequency output signal in dependence on the signal power within the corresponding frequency band.

21. A method according to any one of Claims 18 to 20, wherein the step of processing the audio signal includes the step of generating each frequency output signal in dependence on an average of the signal strength within each frequency band.

22. A method according to any one of Claims 18 to 21, wherein the step of processing the audio signal includes the step of generating each frequency output signal in dependence on a substantially logarithmic function of the signal strength within the corresponding frequency band.

23. A method according to any one of Claims 18 to 22, wherein the frequency bands of the step of analysing the audio signal have centre frequencies forming a substantially exponential series.

24. A method according to any one of Claims 18 to 23, wherein the step of analysing the audio signal includes the step of band-pass filtering the audio signal to select each of the plurality of frequency bands for processing essentially independently of the other frequency bands.

25. A method according to any one of Claims 18 to 24, wherein each modulation output signal represents a signal strength within a frequency band of the corresponding frequency output signal.

26. A method according to Claim 25, wherein the step of processing the frequency output signals includes the step of generating multiple modulation output signals per frequency output signal, each of said multiple modulation output signals being associated with one of a corresponding plurality of frequency bands of the frequency output signal.

27. A method according to Claim 25 or 26, wherein the step of processing the frequency output signals includes the step of generating each modulation output signal in dependence on the signal power within the corresponding frequency band of the corresponding frequency output signal.

28. A method according to any one of Claims 25 to 27, wherein the step of processing the frequency output signals includes the step of generating each modulation output signal in dependence on an average of the signal strength within the corresponding frequency band of the corresponding frequency output signal.

29. A method according to any one of Claims 25 to 28, wherein the step of processing the frequency output signals includes the step of generating each modulation output signal in dependence on a substantially logarithmic function of the signal strength within the corresponding frequency band of the corresponding frequency output signal.

30. A method according to any one of Claims 25 to 29, the frequency bands of the step of analysing the frequency output signals have centre frequencies forming a substantially exponential series.

31. A method according to any one of Claims 25 to 30, wherein the step of analysing the frequency output signals includes the step of band-pass filtering each frequency output signal to select each frequency band for processing essentially independently of any other frequency bands.

32. A method according to any one of Claims 18 to 31, further comprising: inputting the audio signal; processing the modulation output signals to identify an audio signal source likely to be associated with the audio signal; and outputting an identification signal representative of the identified audio signal source.

33. A method according to any one of Claims 18 to 32, further comprising: processing the identification signal, and outputting an alarm signal if the identified audio source relates to an intrusion into a protected space.

34. A carrier medium carrying computer readable code for controlling a computer to carry out the method of any one of Claims 18 to 33.