VEHICLE EVENT ASSESSMENT

Abstract

The disclosure relates to apparatus (300) and automated methods (100, 200) for generating a library of templates (304) corresponding to different known types of motor vehicle event and discriminating between types of event on a motor vehicle. The apparatus (300) comprises the template library (304) and a pattern matching processor (302).

Claims

1. Apparatus for discriminating between types of event on a motor vehicle, comprising: a template library storing a plurality of different templates, each template corresponding to an event type; and a pattern matching processor configured to (i) receive motion sensor data from one or more motion sensors on the motor vehicle, (ii) apply a wavelet transformation to the motion sensor data in order to identify features of transformed motion sensor data, (iii) compare at least one of the identified features of the transformed motion sensor data with templates in the template library and (iv) determine an event type based on the comparison.

2. The apparatus of claim 1 wherein the one or more identified features are coefficients of the transformed motion sensor data, wherein a plurality of coefficients associated with at least one template is provided in the template library, and wherein the pattern matching processor is configured to compare each coefficient of the transformed motion sensor data with the at least one template of a corresponding coefficient provided in the template library.

3. The apparatus of claim 2, wherein the pattern matching processor is configured to match a scale and translation value of each of the coefficients of the transformed motion sensor data with a scale and translation value of the at least one template of the corresponding coefficient provided in the template library.

4. The apparatus of claim 1 comprising the at least one motion sensor.

5. The apparatus of claim 4 comprising only a single type of motion sensor.

6. The apparatus of claim 5 comprising only a single motion sensor.

7. The apparatus of claim 4 wherein the at least one motion sensor comprise one of an accelerometer and a three-dimensional accelerometer.

8. (canceled)

9. The apparatus of claim 7 wherein the accelerometer is configured to be mounted in the vehicle with an axis of acceleration normal to the ground.

10. The apparatus of claim 1 wherein the at least one motion sensor comprise a vibration sensor.

11. (canceled)

12. The apparatus of claim 1 wherein the events are impact events.

13. The apparatus of claim 1 wherein the events are non-impact events.

14. The apparatus of claim 1 wherein the type of event is an acceleration event.

15. The apparatus of claim 1 wherein the pattern matching processor is configured to classify a type of vehicle behaviour based on a number of determinations of acceleration event types, the classification based on the occurrence of each type of acceleration event.

16. The apparatus of claim 13 wherein one or more of the templates in the template library are associated with a risk weighting for the corresponding event type and wherein the classification is also based on the risk weighting of each type of acceleration event.

17. The apparatus of claim 1 wherein at least one of the templates in the template library is each associated with a type of braking event.

18. An automated method for discriminating between types of event on a motor vehicle, comprising: receiving motion sensor data from at least one motion sensor on the motor vehicle; retrieving a plurality of different templates from a template library, each template corresponding to an event type; applying a wavelet transformation to the motion sensor data in order to identify features of transformed motion sensor data; comparing at least one of the identified features of the transformed motion sensor data with the plurality of different templates; and determining an event type based on the comparison.

19. An automated method for generating a library of templates corresponding to different known types of motor vehicle event, comprising: receiving motion sensor data representative of the different types of motor vehicle event; applying a wavelet transformation to the motion sensor data in order to identify features of transformed motion sensor data; for at least some of the different types of motor vehicle event, determining values of at least one indicative feature, each value corresponding with a particular type of motor vehicle event; providing the library of templates comprising the indicative features and an identifier of the particular type of motor vehicle event with which each value corresponds.

20. The automated method of claim 17 wherein the motion sensor data comprises a plurality of examples of each different type of motor vehicle event.

21. The automated method of claim 18 wherein identifying features of the motion sensor data comprises generating a matrix of coefficients using a discrete wavelet transformation, each coefficient having an element associated with one of the plurality of examples of each different type of motor vehicle event.

22. The automated method of claim 19 wherein determining values of the at least one indicative feature comprises performing cluster analysis on the elements of each coefficient and identifying at least one coefficient that provides a separate cluster for each different type of motor vehicle event, and wherein each template comprises a description of a cluster.

23-30. (canceled)

Description

[0053] Embodiments of the invention will now be described, by way of example, with reference to the following figures, in which:

[0054] FIG. 1 illustrates an automated method for generating a library of templates corresponding to different known types of motor vehicle impact event;

[0055] FIG. 2 illustrates an automated method for discriminating between types of impact event on a motor vehicle;

[0056] FIG. 3 illustrates a block diagram representation of an apparatus for discriminating between types of impact event on a motor vehicle and additional associated sensors;

[0057] FIG. 4a illustrates longitudinal accelerometer data for an impact event;

[0058] FIG. 4b illustrates lateral accelerometer data for an impact event;

[0059] FIG. 5a illustrates an impact event record containing sequentially combined longitudinal and lateral accelerometer data from FIGS. 4a and 4b;

[0060] FIG. 5b illustrates three-dimensional data for combining into a single event record;

[0061] FIGS. 6a to 6c illustrate impact records from three different types of impact event;

[0062] FIGS. 6d to 6f illustrate event records from three different types of non-impact event;

[0063] FIG. 7 illustrates a table of motion sensor data points;

[0064] FIG. 8 illustrates a table of coefficients of transformed motion sensor data;

[0065] FIG. 9a illustrates an example of a coefficient that enables three different impact event types to be resolved;

[0066] FIG. 9b illustrates an example of a coefficient that does not enable three different impact event types to be resolved;

[0067] FIG. 10 illustrates a block diagram representation of a portion of a template library; and

[0068] FIG. 11 illustrates a block diagram representation of transformed motion sensor data to be classified.

[0069] The present disclosure relates to using pattern matching to discriminate between a plurality of different types of motor vehicle events, including impact events and non-impact events. Non-impact events include acceleration events, such as braking.

[0070] FIG. 1 illustrates a series of steps of an automated method for generating a library of templates corresponding to different known types of motor vehicle event, which may be impact events. Once generated, the library can be used to categorise sensor data from subsequent impact events, as will be discussed with reference to the method of FIG. 2 and apparatus of FIG. 3. The method of FIG. 1 can be considered to illustrate a learning mode, whereas the method of FIG. 2 illustrates a classification mode. An example according to the method of FIG. 1 is described in further detail with reference to FIGS. 4 to 10. An example according to the method of FIG. 2 is described in further detail with reference to FIGS. 10 and 11. In each of these examples, the method and apparatus are discussed in regard to the analysis of impact events. These systems and apparatus may also be used to classify non-impact events, such as acceleration/braking events.

[0071] As an initial step in the method 100 of FIG. 1, motion sensor data representative of different types of motor vehicle impact event are received at step 102. The motion sensor data can take the form of a digitised waveform, or waveforms. The motion sensor data may include accelerometer data or vibration meter data obtained from motion sensors situated on or inside a motor vehicle, such as a road vehicle, watercraft or aircraft. It is desirable to ensure that the motion sensors acquire the motion sensor data at a sufficient frequency to comply with the Nyquist limit of the waveform.

[0072] Unprocessed, or unfiltered, motion sensor data typically comprises real information, system noise, environmental noise, sampling errors and signal aliasing. Filtering can be applied to the motion sensor data to remove artefacts. A 4th order Butterworth filter has been found to be suitable to apply to raw motion sensor data for some applications. The Butterworth filter can be applied bi-directionally to remove phase errors (for example, see filter type CFC 60 featured in SAE J211: Instrumentation for Impact Test, Part 1, Electronic Instrumentation).

[0073] Motion sensor data from a particular impact event may be provided as an impact record. In order to train the system, a plurality of examples of each different type of motor vehicle impact event may be received to form a training set.

[0074] Once the training set of impact data has been received (step 102) and, optionally, processed, features of the motion sensor data can be identified at step 104. Identifying the features can be achieved using a variety of known signal processing techniques. Identifiable features for discriminating impact type may relate to variations in the waveform of the motion sensor data (in the time domain). For example, identifiable features for discriminating between different impact types may include peaks and troughs in waveform data. The identifiable features have a number of values, such as the duration of the impact event, magnitude, frequency, relative direction, and envelope size. A plurality of features may be considered in order for pattern matching to be applied, as opposed to a single comparison with a threshold level.

[0075] Alternatively, a transformation may be applied to the waveform of the motion sensor data in order to identify features for discriminating impact type. A Fourier transform may be applied to the received waveforms in order to determine frequency information. In this case, one or more frequency components from the transformed data may be identifiable features for discriminating impact type. A disadvantage of applying a Fourier transform to an entire motion sensor data record is that temporal information, which can be useful in categorising impact types, is lost. In order to address this problem, each dataset can be broken down into windowed periods of time and each window can be individually Fourier transformed. However, such a technique may require prior knowledge of the expected duration of each window for best results.

[0076] Applying a wavelet transformation to the waveform of the motion sensor data provides transformed motion sensor data that includes temporal information as well as information regarding the amplitude/power and frequency components of the motion sensor data. Where the waveform is provided as a digitised signal, a discrete wavelet transformation can be applied in order to generate a matrix of coefficients of sensor data corresponding to each of the motor vehicle impact event records. The coefficients relate to identified features of the motion sensor data. The matrix can be considered as a table of elements with columns of coefficients and rows of impact records. Each element of the matrix is therefore associated with both a particular coefficient and a particular impact event. Each coefficient is associated with a number of elements. Each impact record is also associated with a number of elements.

[0077] Alternatively, where the waveform is provided as a continuous signal, a continuous wavelet transformation may be applied to the waveform. The transformed waveform may be considered to provide the identified features of the motion sensor data in such an example.

[0078] Identifiable features that have a value that varies depending upon the type of impact event can be considered to be indicative features. For at least some of the different types of motor vehicle impact event, values of one or more indicative features that correspond with a particular type of motor vehicle impact event are determined at step 106.

[0079] Where wavelet transformation has been used to provide a matrix of coefficients of sensor data, cluster analysis can be performed on the matrix of coefficients in order to identify features that provide a separate cluster for each different type of motor vehicle impact event. Coefficients with elements that take distinguishable values depending on the impact type of impact to which the elements relate can be considered to be indicative features. In this way, the values of the one or more indicative features that are associated with a particular type of motor vehicle impact event can be determined.

[0080] The values of the one or more indicative features are provided, together with an identifier of the particular type (or, in some examples, types) of motor vehicle impact event with which each value corresponds, as a template library (at step 108). Each value of an indicative feature associated with one particular type of motor vehicle impact event can be considered to be a template for that particular type of motor vehicle impact.

[0081] The template library is stored for subsequent use. The library can be stored centrally and/or provided on a mobile device, such as an in-vehicle device, for subsequent classification of impact events.

[0082] FIG. 2 illustrates an automated method 200 that may be used after an incident or event has occurred in order to discriminate between, or classify, different types of impact or non-impact event on a motor vehicle, using the template library. The monitoring of non-impact events during use of the vehicle enables a number of events to be monitored so that the behaviour of the vehicle can be classified. The method 200 is described in further detail below with regard to an impact event example.

[0083] The method 200 comprises the step of receiving incident motion sensor data from only one, or more than one, motion sensor on the motor vehicle (at step 202). The sensors are typically of the same kind as those used to produce the template library at step 102. The accelerometer data may be monitored continuously while the vehicle is in motion. Alternatively, accelerometer data may be monitored only when the acceleration along an axis is above a pre-set threshold, or when a predetermined geographical area is entered. For example, monitoring may be triggered in response to a positioning system (such as a satellite based positioning system such as the global positioning system, GPS) indicating that a pre-determined area has been entered. In this way, data may be actively monitored around accident or crime hotspots.

[0084] A plurality of different templates are retrieved from the template library at step 204. Each template corresponds to an impact type. The retrieval of the templates at step 204 can take place before, after or simultaneously with the receipt of motion sensor data at step 202.

[0085] Features of the incident motion sensor data are identified in a similar manner to that used in step 104 when generating the template library. The features may provide a quantitative measure of the one-, two, or three-dimensional acceleration profile of a vehicle over a predetermined period of time. The features may encompass absolute or relative changes in acceleration along an axis, or a periodicity of an acceleration change, for example. The features of the incident motion sensor data are compared with each of the plurality of templates at step 206. An impact type is determined based on the outcome of the comparisons between the incident motion sensor data and each of the plurality of templates. The impact event type may be determined based on an identifier of the particular impact event type with which a template that corresponds best with the features of the incident motion sensor data is associated.

[0086] A number of attributed distinguish braking events form impact events. For example, impact events may occur over a time period of about 60 ms or 70 ms, in some examples, and typically less than 250 ms, irrespective of the type of impact. Pothole impacts may occur in a similar time frame to more serious impacts. Braking and acceleration events may have an acceleration in the range of 1 g to 1.3 g or 1.5 g. Impact events may have a maximum acceleration that is greater than 2 g. Braking events related to non-impact events may occur over a time period of greater than a 0.5 seconds or a second.

[0087] Non-impact events may not result in the vehicle coming to rest whereas a vehicle will typically stop after an impact event has occurred.

[0088] In the case where the method is used to classify non-impact events, the steps 202 to 208 of the method 200 may be repeated so that the types of event associated with a plurality of events are determined. For example, the motion sensor data may be associated with a plurality of different events that occur while the vehicle is being driven. The motion sensor data may be compared with templates in the template library in order to determine an event type associated with each of the plurality of events based on the comparisons.

[0089] The method may further comprise the step of classifying a type of vehicle behaviour based on a number of determinations of acceleration event types. For example, the method may comprise recording each time that a type of event occurs. Each type of event may be assigned a risk weighting, which can take a numeric value. Each template may be associated with a risk weighting for the corresponding event type. A statistical profile of the behaviour of the vehicle can be determined based on the occurrence of each type of acceleration event and the risk associated with each type of event. For example, cadence braking may be assigned a higher value than distracted braking, which in turn may have a substantially higher value than observant type braking. In this way, the behaviour of a driver of a vehicle over a driving session, or other period of time, can be reflected by an aggregate score. The statistical profile, or score, may be recorded for each driving session. Meta-data such as the time of day, identity of the vehicle driver or location of the vehicle may be associated with the statistical profile. Data concerning each session can be sent to a remote computer for storage or further analysis.

[0090] FIG. 3 illustrates an apparatus 300 for discriminating between types of impact event on a motor vehicle using the method 200 of FIG. 2. The same, or a similar, apparatus 300 can also be used to generate the template library by performing the method 100 of FIG. 1. Although the example discussed below relates to impact event classification, the apparatus 100 may also be used to discriminate between types of non-impact event on a motor vehicle or generate the template library for non-impact events.

[0091] The apparatus may be provided by a mobile device, such as a mobile telephone or mobile computer, a vehicle-mounted device, or a server remote from the vehicle. Alternatively, the apparatus may be distributed between a local device, provided in a vehicle, and a server remote from the vehicle.

[0092] The apparatus 300 comprises a pattern matching processor 302 and a template library 304. The pattern matching processor 302 is configured to receive motion sensor data from one or more motion sensors 306, 308 on the motor vehicle. The motion sensors in this example include an accelerometer 306 and a vibration sensor 308. A one-, two- or three-dimension accelerometer may be used, depending on the application. The accelerometer may be a micro-electromechanical systems MEMS accelerometer. A one-dimensional accelerometer may be suitable for applications where the processing power of the pattern matching processor 302 is constrained. Alternatively, other types of sensors could be used, such as an inertial measurement unit (IMU). An IMU can be provided with one, two or three axis rotation data. The motion sensors 306, 308 are preferably integrated with the vehicle to reduce damping of measured movement of the vehicle. The apparatus 300 may also comprise the motion sensors 306, 308, for example, where the apparatus 300 is provided as a vehicle mounted device. Examples of the apparatus 300 may be implemented using only a single sensor and still provide sufficient data in order to classify a plurality of different types of event because of the techniques implemented by the pattern matching processor 302. Using only a single sensor, or only a single type of sensor, can result in a simplified apparatus 300.

[0093] Motion sensor data related to an impact event can be obtained by buffering motion sensor data obtained from the one or more motion sensors 306, 308 and then automatically capturing information from the buffer for a period surrounding when an impact event is observed. Alternatively, data can be continuously recorded for later analysis. As a further alternative, the apparatus 300 can be configured to continuously analyse blocks of motion sensor data obtained from one or more motion sensors 306, 308.

[0094] The template library 304 contains a plurality of different templates, each template corresponding to an impact type. The pattern matching processor 302 is configured to compare the motion sensor data with templates in the template library (at step 206) and determine an impact type based on the comparison (at step 208). The template library 304 is stored in a memory of the apparatus 300, which may be provided remote from the vehicle. However, in order for the pattern matching processor 302 to process the templates there is typically at least transitory storage of the template library 304 in the same apparatus in which the pattern matching processor 302 is housed.

[0095] The pattern matching process 302 and template library 304 may be used to compare non-impact event data instead of, or as well as, impact event data. The pattern matching processor 300 may be further configured to classify a type of vehicle behaviour based on a number of determinations of acceleration event types. Alternatively, the determinations of the accelerations event types may be sent to a remote computer and the remote computer may classify a type of vehicle behaviour based on a number of determinations of acceleration event types.

[0096] The apparatus 300 may comprise a power supply that is configured to receive power from the vehicle when in use such as from the vehicle battery. The received power supply may be 12 V or 24 V.

[0097] A preferred form of pattern matching for the methods of FIGS. 1 and 2 uses wavelet transformation methods. Wavelet transformations retain both frequency and location information from a data source. As such, the use of a wavelet transformation instead of a Fourier transformation can be advantageous because the wavelet transformation provides temporal resolution. The application of a discrete wavelet transformation method is described further with regard to FIGS. 4 to 10.

[0098] FIGS. 4a and 4b illustrate longitudinal accelerometer data 402 and lateral accelerometer data 404 for a single impact event as a function of time. This accelerometer data is an example of motion sensor data that can be obtained from a 2D or 3D accelerometer. The longitudinal accelerometer data 402 provides a waveform with a first peak 406 and a second peak 408. The first peak 406 has a maximum amplitude of around 0.1 to 0.2 g.

[0099] The start of the second peak 408 follows the first peak by about 50 ms. The second peak 408 has a full width at half maxima of about 80 ms and a maximum amplitude of +2.5 g. The lateral accelerometer data 404 provides a waveform with a third peak 410 and a series of peaks 412. The third peak 410 occurs at a similar time as the first peak 406 and has a maximum amplitude of around 0.1 to 0.2 g. The series of peaks 412 occurs at a similar time to the second peak 408 and has a maximum amplitude of around 0.1 to 0.2 g.

[0100] In order to generate a template for the type, T, of impact event to which the motion sensor data relates, the longitudinal and lateral accelerometer data 402, 404 are combined into a single event record, R.sub.1, which relates to that particular impact event. In general, an event record, R.sub.i, contains data from each sensor that recorded a particular impact event.

[0101] FIG. 5a illustrates a sequentially combined impact event record R.sub.1 comprising the longitudinal and lateral accelerometer data 402, 404 of FIGS. 4a and 4b. The data is shown on axes of acceleration against sample number. The sample number is related to time (it is proportional to time in this example) within each of the longitudinal and lateral accelerometer data 402, 404 portions of the event record R.sub.1.

[0102] In a similar way, in order to generate a template for a type of a non-impact acceleration event to which the motion sensor data relates, the x, y and z dimensions of three dimensional accelerometer data can be combined into a single acceleration event record which relates to that particular acceleration event. In general, an event record contains data from each sensor that recorded a particular acceleration event.

[0103] FIG. 5b illustrates x, y and z dimensions of accelerometer data 522, 524, 526 for a single acceleration event as a function of time. The acceleration event relates to a vehicle going over a speed bump at an impact instant 528. A one-dimensional accelerometer may be provided in order to detect speed bump acceleration events. A one-dimensional (or higher dimension) accelerometer may be mounted in a vehicle with an axis of acceleration normal to the ground in order to obtain z-axis acceleration events.

[0104] Lateral acceleration occurs in a direction that is transverse, or normal to the axis of the vehicle. Additional behaviour, such as swerving which includes a lateral component, can be identified by analysis of two- or three-dimensional acceleration data.

[0105] Analysis of acceleration normal to the ground (z-axis acceleration) can be used to classify risky behaviour such as the vehicle driving over a speed bump or over a bridge or crest in a road (which may be associated with a loss of control) at an excessive speed at a relatively low temporal resolution. This accelerometer data in FIG. 5b is provided by a 3D accelerometer operating at 100 Hz, which is a similar resolution to some conventional accelerometers. A lower resolution of 10 or 20 Hz may also be used in some cases and still provide data sufficient for classifying non-impact events. By reducing the temporal resolution of the accelerometer or reducing the number of dimensions of accelerometer data, the power consumption of the accelerometer and data processing apparatus can be reduced. However, a higher rate may still provide more information that is useful in some applications. The example rate may be chosen to ensure that sufficient detail is resolvable without the undue introduction of artefacts.

[0106] Many more specific types of behaviour, including braking behaviour, can be distinguished between by analysing a one-dimensional longitudinal acceleration profile (x-axis acceleration, in the direction of an axis of the vehicle), such as those illustrated in FIGS. 6d to 6f.

[0107] FIGS. 6a to 6c relate to example event records (also known as impact records) for impact events (using the type of impact event sensor data discussed above in FIG. 5a) and FIGS. 6d to 6f relate to example event records for non-impact events (using the type of non-impact event sensor data discussed above in FIG. 5b).

[0108] FIGS. 6a to 6c illustrate respective impact records R.sub.1, R.sub.21, R.sub.31 from three different types T.sub.1, T.sub.2, T.sub.3 of real life impact event. As in FIG. 5, the data is shown on axes of acceleration against sample number. In this case, the type T of each impact is known so the data can be used to train a pattern matching processor, or other classifier.

[0109] For illustrative purposes, the impact record R.sub.1 exemplifies a full width front collision (impact event type T.sub.1). The impact record R.sub.21 exemplifies a front offset collision (impact event type T.sub.2). The impact record R.sub.31 exemplifies an offset rear collision (impact event type T.sub.3).

[0110] For each type T of impact event, a number of different records, R.sub.i, are obtained in order to produce a data structure, or matrix of data points, as discussed below with reference to FIG. 7.

[0111] FIGS. 6d to 6e illustrate respective event records P.sub.1, P.sub.21, P.sub.31 from three different types of real life acceleration events. Event records are referred to as P records in FIGS. 6d to 6f in order to avoid confusion with the impact records R discussed with regard to the impact classifying examples of FIGS. 6a to 6c.

[0112] Drivers may use different types of braking to deal with different types of situations. For example, types of braking employed by drivers include cadence-type, distracted-type and observant-type braking.

[0113] For illustrative purposes, the event record P.sub.1 in FIG. 6d exemplifies a cadence-type braking event (acceleration event type). Cadence-type braking may occur when an impatient driver follows another vehicle too closely. The driver of the following vehicle generates a series of cycles 602 of positive acceleration 604 followed by braking 406 as they attempt to pressure the vehicle in front to pull over. The duration of each acceleration and braking cycle is about 0.75 seconds in this example.

[0114] The event record P.sub.21 in FIG. 6e exemplifies a distracted-type braking event (acceleration event type). Distracted-type braking may occur when a driver is either paying insufficient attention to the road ahead or is actively distracted, by a mobile phone for example. Distracted-type braking may be characterised by a panicked, sharp stabbing 610 on brakes possibly followed by a prolonged period 612 of braking. The duration of the stab 610 on the brakes is less than half a second and the prolonged period 612 of braking is around 2 seconds in duration in this example

[0115] The event record P.sub.31 in FIG. 6f exemplifies an observant-type braking event (acceleration event type). Observant-type braking may occur when a driver who is fully aware of a situation observes something unexpected happening ahead of them, such as a car suddenly pull out or a pedestrian step into the road. A steady pattern of braking may be applied in a measured way which brings the vehicle to a required reduced speed or a safe stop. The steady braking lasts for around 2.5 seconds in this example.

[0116] A block diagram of an example resultant data structure 700 of the training set discussed with regard to the impact data examples in FIGS. 6a to 6c is illustrated in FIG. 7. A similar data structure may be provided for the example non-impact data described with regard to FIGS. 6d to 6f. The training set contains three different impact event types T1, T2, T3. Each impact event type contains 10 examples (n=10) of impact records (e.g. R.sub.1, R.sub.2, R.sub.N). Each impact record R.sub.1-R.sub.3a contains 802 samples (z=802), or data points, z.sub.i.

[0117] Feature extraction, or identification is performed by applying a discrete wavelet transformation to each impact record R using a 4 level Haar wavelet. Such functions are available in standard applied mathematics packages, such as Matlab.

[0118] FIG. 8 illustrates a table 800, or matrix, of coefficients of transformed motion sensor data. A coefficient is an example of an identified feature provided by a wavelet transformation. The transformed matrix contains 802 coefficients, a.sub.1-a.sub.802, one for each data point z.sub.i. Each coefficient has 30 elements, one element, z.sub.i, for each impact record, R.sub.1-R.sub.3n. Each coefficient a.sub.1-a.sub.802 characterises the waveform shapes of the impact records R.sub.1-R.sub.3n at different scales and translations. Scale relates to the frequency of the original waveform and translation contains relative, rather than absolute, temporal information about the waveform.

[0119] Indicative features of the motion sensor data are determined by selecting a number of indicative coefficients, which are an example of an indicative feature for the wavelet transformation example. Two of these indicative coefficients, a.sub.6 and a.sub.11 are highlighted in FIG. 8. An indicative coefficient (e.g. a.sub.6, a.sub.11) is a coefficient with elements z.sub.i that are distinguishable from one another based on the type T of impact event to which the elements z.sub.i relate. That is, the value of an element is indicative of a particular type of impact event. The ability of each coefficient a.sub.1-a.sub.802 to provide effective clustering of different types T of impact event across all of the impact records R.sub.1-R.sub.3, and therefore differentiate between the different types T of impact event, can be calculated using a variety of statistical methods including cluster analysis.

[0120] In this example, 20 of the 802 coefficients were selected as indicative coefficients, or indicative features, using the KMeans clustering method. These 20 coefficients (e.g. a.sub.6-a.sub.11) are those that produce the highest separation between centroids of groups of elements z that relate to the three different types T.sub.1-T.sub.3 of impact event. The centroid and area of these groups can be calculated using 2 Standard Deviations to rule out outliers. These indicative coefficients a.sub.6, a.sub.11 are selected to form a subset of coefficients a.sub.6, a.sub.11.

[0121] Each of the indicative coefficients can be plotted against each other with each plot containing 30 data points representing the 30 impact records R.sub.1-R.sub.3n within the training set.

[0122] FIG. 9a illustrates a plot of an indicative coefficient a.sub.6 that enables the three different impact event types T.sub.1-T.sub.3 to be resolved. The data plotted in FIGS. 9a and 9b relates to different types of impact events of a vehicle, rather than braking events, but the principle of analysis described below is not dependent on the type of motor vehicle event. The elements z.sub.6 of the coefficient a.sub.6 have a scale value and a translation value. The elements z.sub.6 also have an amplitude, or power, associated with them, but it is not necessary in all cases to consider this amplitude when determining the clustering of coefficients. In practice the amplitude may be dependent on the location of the sensing device, hence the overall form (and thus the centroid weight) may be a simpler and more reliable value in some applications.

[0123] The elements Zs are segregated into three groups 902, 904, 906. Each group 902, 904, 906 is associated with a different impact event type T.sub.1, T.sub.2, T.sub.3. The separation of the elements z.sub.6 into distinguishable groups means that this coefficient may provide templates for the known impact event types T.sub.1, T.sub.2, T.sub.3.

[0124] FIG. 9b illustrates, for comparison with FIG. 9a, a plot of a coefficient a.sub.5 that does not enable the three different impact event types T.sub.1-T.sub.3 to be resolved. As in FIG. 9a, the elements z.sub.5 of the coefficient a.sub.5 are displayed on an axis of scale against an axis of translation. The elements z.sub.5 are segregated into only two groups 908, 910. The elements in a first group 908 are associated with a first type T.sub.1 of impact event. The elements in a second group 910 are associated with either a second type T.sub.2 or a third type T.sub.3 of impact event. The inability to distinguish between the second type T.sub.2 and third type T.sub.3 of impact event means that this coefficient cannot be used to provide templates that distinguish the second type T.sub.2 and third type T.sub.3 of impact event.

[0125] Having determined a set of indicative coefficients a.sub.6, a.sub.11 that effectively separate the impact records R.sub.1-R.sub.3n into discrete clusters of values related to the respective impact types, these indicative features are then stored as learnt data, or templates, for use in a classification mode. Specifically, the centroid, or another statistical description, of each group of elements is a value that corresponds to a particular impact type. Each value is an example of a template. These templates are stored together with an identifier of the particular type of motor vehicle impact event associated with which each template, as a template library.

[0126] By providing a single labelled example of each event record R.sub.1-R.sub.3, a classification can easily be converted into a more reader friendly output as descriptive text strings i.e. Full-Width Crash Front for the impact event type T.sub.1.

[0127] FIG. 10 illustrates a portion of a template library 1000. The template library 1000 comprises each of the coefficients of FIG. 8 identified as being indicative of the type of impact events. Each coefficient a.sub.6, a.sub.11 is associated with a number of templates t.sub.1,6-t.sub.3,11. The templates correspond to the statistical description identified as corresponding to the respective types of impact event T.sub.1, T.sub.2, T.sub.3. As such, each descriptor t.sub.1,6-t.sub.3,11 is a value of the coefficient that is associated with a particular impact event type T.sub.1, T.sub.2, T.sub.3. In this example, the template library also contains a label associated with each impact event type T.sub.1, T.sub.2, T.sub.3. The relationships between each of the classes of data considered in relation to the wavelet transformation embodiments are described below.

TABLE-US-00001 Table of relationships Class (space) Description Example Impact type, T A type of impact full width front collision, T.sub.1; (label) front offset collision, T.sub.2; offset rear collision, T.sub.3 Impact Relates to a specific impact event of an R.sub.1, R.sub.2, R.sub.n, R.sub.21, R.sub.22, R.sub.2n, Record, R impact type T R.sub.31, R.sub.32, R.sub.3n (label) Data point, z Motion sensor data associated with an Each of z.sub.1-z.sub.802 (amplitude, impact record R contains data points time) element, z Transformed motion sensor data Each of z.sub.1-z.sub.802 (scale, associated with an impact record R translation, contains elements. Each element is amplitude) associated with a coefficient a.sub.1-a.sub.802, which acts as an element index. Indicative Coefficients that can be used to a.sub.6, a.sub.11 feature/ distinguish between impact types T. Indicative Associated with elements z that can be coefficient separated into different groups (index) depending on the impact type T of the record R that an element z belongs to. Template A value of a coefficient that is indicative t.sub.1, 6, t.sub.2, 6, t.sub.3, 6, t.sub.1, 11, t.sub.2, 11, (scale, of a particular impact type. Can be a t.sub.3, 11; centroid of 902, 904, translation) descriptor of a group of elements of an 906 indicative coefficient.

[0128] In the classification mode, the method of FIG. 2 or apparatus of FIG. 3 is used to classify received motion sensor data using the template library 1000. The use of a discrete wavelet transformation in such a method or apparatus is discussed below.

[0129] FIG. 11 illustrates a transformed impact record 1100 to be classified. Such motion sensor data is related to an unknown type of impact event and is received motion sensor data from one or more sensor 306, 308 on a motor vehicle (step 202 of method 200).

[0130] The impact record R.sub.sample is transformed into a number of coefficients (the transformed impact record 1100) using the same or similar discrete wavelet transformation function used to generate the template library. In the classification mode, there is typically only one event record R.sub.sample and so each coefficient a.sub.1-a.sub.802 has a single element z.sub.1-z.sub.802.

[0131] The plurality of templates are retrieved 204 from the template library 1000. The coefficients as, all of the impact record 1100 that correspond to the coefficients of the templates stored in the template library 1000 are selected. In this way, features of the motion sensor data are identified. The elements z.sub.6, z.sub.11 of the selected coefficient a.sub.6, a.sub.11 of the impact record 1100 are compared to the templates t.sub.1,6-t.sub.3, 11 (at step 206). Both the elements z.sub.6, z.sub.11 and the templates t.sub.1,6-t.sub.3, 11 have a scale and translation associated with them because the elements are wavelet transformed values and the templates t.sub.1,6-t.sub.3,11 relate to the centroid of groups of elements as discussed with reference to FIGS. 8 to 10.

[0132] The comparison between the elements z.sub.6, z.sub.11 and templates t.sub.1, 6-t.sub.3, 11 can be performed in a number of ways. For example, a degree of correlation between an element (e.g. z.sub.6) of a coefficient a.sub.6 of the impact record 1100 and each of the templates t.sub.1,6, t.sub.2,6, t.sub.3,6 that belong to the corresponding coefficient a.sub.6 of the template library 1000 can be generated. From this comparison it can be determined if the element z.sub.6 matches any of the templates t.sub.1,6, t.sub.2,6, t.sub.3,6. That is, whether the element z.sub.6 fall within a sufficiently close proximity of a template in the scale and translation space. Alternatively, it can be determined which template the element z.sub.6 is closest to in the scale-transformation space. This process may be repeated for each of the selected coefficients z.sub.6, z.sub.11. The determined impact type may therefore relate to the impact type T associated with the majority of templates that match the elements z.sub.6, z.sub.11 of the selected coefficient a.sub.6, a.sub.11.

[0133] An impact type classification can be converted into a more reader friendly output as descriptive text strings. That is the label [e.g. full width front crash] associated with the determined impact type T.sub.1 can be provided as an output. A degree of correlation between the matching templates and the impact record R.sub.sample, or another confidence ranking to indicate the quality of fit across the selected coefficients, can also be provided as an output.