Vehicle classification system using a passive audio input to a neural network

• US 5,619,616 A
• Filed: 04/25/1994
• Issued: 04/08/1997
• Est. Priority Date: 04/25/1994
• Status: Expired due to Fees

• Current Assignee: 3M Company
• Original Assignee: 3M Company
• Inventors: Hamerly, Michael E., Brady, Mark J.
• Primary Examiner(s): Hafiz, Tariq R.
• Application Number:
  US08/232,758
• Time in Patent Office: 1,079 Days
• Field of Search: 395/20-25, 395/27, 395/2, 395/2.41, 395/2.4, 382/155-159, 367/131, 364/434, 381/41-43
• US Class Current: 706/20
• CPC Class Codes: G01H 3/08 Analysing frequencies present in complex vibrations, e.g. comparing harmonics present acoustic presence detection G01V1/001
• CPC Class Codes: G01V 1/001 Acoustic presence detection
• CPC Class Codes: G06F 18/2414 Smoothing the distance, e.g. radial basis function networks [RBFN]
• CPC Class Codes: G06F 2218/10 by analysing the shape of a waveform, e.g. extracting parameters relating to peaks
• CPC Class Codes: G08G 1/015 with provision for distinguishing between two or more types of vehicles, e.g. between motor-cars and cycles

Abstract

A system for classifying vehicles based on the sound waved produced by the vehicles receives analog sound pressure levels and converts them to a power spectrum. Fuzzification functions, such as asymmetric wedge shaped functions, are convoluted with the power spectrum to create a vector that characterizes the power spectrum while reducing the dimensionality of the characterizing vector. A neural network analyzes the characterizing vector and produces a classification designator indicative of the class of the object associated with the analog sound pressure levels received by the system.

Claims

1. A passive audio classification system for classifying an object emitting sound, said system comprising:

an analog to digital converter for converting analog sound wave characteristics associated with said object to digital sound wave characteristics;

means for converting said digital sound wave characteristics, said digital sound wave characteristics measured from a time interval, into a power spectrum;

means for applying a fuzzification function to said power spectrum to create a vector of a predetermined dimension, said vector characterizing said power spectrum; and

a neural network for receiving said vector and for producing a classification designator based on said vector.

2. The passive audio classification system according to claim 1, further comprising processor means for analyzing classification designators produced by said neural network over a plurality of said time intervals and for generating a signal indicative of the class of said object.

3. The passive audio classification system according to claim 1 further comprising measuring means for measuring analog sound wave characteristics of said sound emitted from said object.

4. The passive audio classification system according to claim 3, wherein said measuring means is a microphone.

5. The passive audio classification system according to claim 1, wherein said means for converting said digital sound wave characteristics minimizes peak reduction when converting said digital sound wave characteristics to said power spectrum.

6. The passive audio classification system according to claim 1, wherein said means for converting said digital sound wave characteristics only outputs predominant frequencies in the power spectrum.

7. The passive audio classification system according to claim 6, wherein said means for converting said digital sound wave characteristics outputs a power value of frequencies where said power value is at a local maximum.

8. The passive audio classification system according to claim 6, wherein said means for converting said digital sound wave characteristics outputs a power value of frequencies when said power values are one of a predetermined number of highest local maxima.

9. The passive audio classification system according to claim 1, wherein said means for applying said fuzzification function applies a predetermined number of overlapping asymmetric wedge shaped functions to said power spectrum to create said predetermined dimensional vector.

10. The passive audio classification system according to claim 9, wherein each of said overlapping asymmetric wedge shaped functions has a unit height, is centered at a frequency of 2ⁿ and is defined between the frequencies of 2^n-1 and 2ⁿ as:

##EQU5## and is defined between the frequencies of 2ⁿ and 2ⁿ⁺¹ as;

##EQU6## where x is the frequency and y is the height of said function.

11. The passive audio classification system according to claim 1, wherein said analog sound wave characteristics associated with said object are sound pressure levels.

12. The passive audio classification system according to claim 1, wherein said means for converting said digital sound wave characteristics comprises a root mean square converter.

13. The passive audio classification system according to claim 1, wherein said means for applying a fuzzification function to said power spectrum comprises a finite impulse response filter.

22. The passive audio classification system according to claim 1 wherein the objects that are classified are vehicles.

14. A method of classifying an object emitting sound, said method comprising the steps of:

receiving analog sound wave characteristics associated with said object, said analog sound wave characteristics measured from a time interval;

convening said analog sound wave characteristics to digital sound wave characteristics;

converting said digital sound wave characteristics to a power spectrum;

applying a fuzzification function to said power spectrum to create a vector of a predetermined dimension for characterizing said power spectrum; and

producing a class designator at a neural network based on said vector.

15. The method of classifying an object emitting sound according to claim 14, further comprising the steps of analyzing classification designators produced by said neural network for a plurality of said time intervals and generating a signal indicative of the class of said object.

16. The method of classifying an object emitting sound according to claim 15, wherein the step of analyzing classification designators from a plurality of said time intervals comprises the steps of:

comparing the number of class designators in each different class; and

determining which said class has the greatest number of class designators.

17. The method of classifying an object emitting sound according to claim 14, further comprising the step of minimizing peak reduction when converting digital sound wave characteristics to said power spectrum.

18. The method of classifying an object emitting sound according to claim 17, wherein minimizing peak reduction comprises the steps of:

dividing a frequency bin into a predetermined number of frequency intervals;

converting said digital sound wave characteristics at each of said predetermined frequency intervals to said power spectrum, each converted frequency having a central peak;

analyzing said series of converted signals to determine which converted signal has the highest central peak; and

assigning a power value from said signal with the highest central peak for said power value of said frequency bin.

19. The method of classifying an object emitting sound according to claim 17, wherein the step of dividing a frequency bin comprises the step of dividing said frequency bin into intervals of 0.1 Hz.

20. The method of classifying an object emitting sound according to claim 14, wherein applying fuzzification functions to said power spectrum comprises the steps of:

convoluting each of a plurality of fuzzification functions over a range of frequencies to produce a convoluted result, each said function centered at a characterizing frequency; and

assigning each said convoluted result to a corresponding component of said vector.

21. The method of classifying an object emitting sound according to claim 20 wherein said characterizing frequency of each said function is 2ⁿ and wherein said function is defined between the frequencies of 2^n-1 and 2ⁿ as:

##EQU7## and is defined between the frequencies of 2ⁿ and 2ⁿ⁺¹ as;

##EQU8## where x is the frequency and y is the height of said function.