Neural network radar processor

• US 6,366,236 B1
• Filed: 08/11/2000
• Issued: 04/02/2002
• Est. Priority Date: 08/12/1999
• Status: Expired due to Fees

• Current Assignee: Automotive Systems Laboratory, Incorporated (Ningbo Junsheng Investment Group Co., Ltd.)
• Original Assignee: Automotive Systems Laboratory, Incorporated (Ningbo Junsheng Investment Group Co., Ltd.)
• Inventors: Farmer, Michael E., Jacobs, Craig S., Cong, Shan
• Primary Examiner(s): Blum, Theodore M.
• Application Number:
  US09/637,044
• Time in Patent Office: 599 Days
• Field of Search: 342/195, 342/118, 342/200
• US Class Current: 342/195
• CPC Class Codes: B60R 21/0134 responsive to imminent contact with an obstacle , e.g. using radar systems
• CPC Class Codes: G01S 13/34 using transmission of continuous, frequency-modulated waves while heterodyning the received signal, or a signal derived therefrom, with a locally-generated signal related to the contemporaneously transmitted signal
• CPC Class Codes: G01S 13/584 adapted for simultaneous range and velocity measurements
• CPC Class Codes: G01S 13/931 of land vehicles
• CPC Class Codes: G01S 7/417 involving the use of neural networks

Abstract

A neural network radar processor (10) comprises a multilayer perceptron neural network (100.1) comprising an input layer (102), a second layer (122), and at least a third layer (124), wherein each layer has a plurality of nodes (108), and respective subsets of nodes (108) of the second (122) and third (124) layers are interconnected so as to form mutually exclusive subnetworks (120). In-phase and quadrature phase time series from a sampled down-converted FMCW radar signal (19) are applied to the input layer, and the neural network (100) is trained so that the nodes of the output layer (106) are responsive to targets in corresponding range cells, and different subnetworks (120) are responsive to respectively different non-overlapping sets of target ranges. The neural network is trained with signals that are germane to an FMCW radar, including a wide range of target scenarios as well as leakage signals, DC bias signals, and background clutter signals.

Claims

1. A neural network radar processor, comprising:

a. an input layer comprising a first plurality of nodes, wherein each node of said first plurality of nodes has at least one input and an output, an input from each of said first plurality of nodes is operatively connected to a sample of a time series of radar data, and each node of said first plurality of nodes is operatively connected to a different time sample;

b. a second layer comprising a second plurality of nodes, wherein each node of said second plurality of nodes has a plurality of inputs and an output, each output from said first plurality of nodes is operatively connected to an input of each of said second plurality of nodes, and each input of said plurality of inputs of a node of said second plurality of nodes is operatively connected to a different node of said first plurality of nodes;

c. a third layer comprising a third plurality of nodes, wherein each node of said third plurality of nodes has a plurality of inputs and an output, said third plurality is equal in number to said second plurality, said second plurality of nodes comprises a fourth plurality of mutually exclusive subsets of nodes, said third plurality of nodes comprises a fifth plurality of mutually exclusive subsets of nodes, said fifth plurality is equal in number to said fourth plurality, and any input to a node of one of said fifth plurality of subsets is operatively connected only to an output of a node of one of said fourth plurality of subsets; and

d. an output layer comprising a sixth plurality of nodes, wherein each node of said sixth plurality of nodes has a plurality of inputs and an output, said sixth plurality of nodes comprises a seventh plurality of mutually exclusive subsets of nodes, said seventh plurality is equal in number to said fifth plurality, any node of one of said seventh plurality of subsets is operatively connected only to a node of one of said fifth plurality of subsets, an output of a node of said sixth plurality of nodes provides an indication of a target in a particular range cell, outputs of nodes of one of said seventh plurality of subsets provide indications of targets in range cells for a first set of target ranges, and outputs of nodes of another of said seventh plurality of subsets provide indications of targets in range cells for a second set of target ranges, whereby said first and second sets of target ranges are not overlapping.

2. A neural network radar processor as recited in claim 1, wherein said first plurality of nodes comprise first and second subsets of nodes, an input from each of said first subset of nodes is operatively connected to a sample of a first time series of radar data, each node of said first subset is operatively connected to a different time sample, an input from each of said second subset of nodes is operatively connected to a sample of a second time series of radar data, said second time series is of quadrature phase to said first time series, each node of said second subset is operatively connected to a different time sample, and said first and second subsets of said first plurality of nodes correspond in time.

3. A neural network radar processor as recited in claim 1, wherein said output layer is the same as said third layer, and said sixth plurality of nodes are the same nodes as said third plurality of nodes.

4. A neural network radar processor as recited in claim 1, wherein said output layer comprises a layer different from said third layer, and said sixth plurality of nodes are distinct from said third plurality of nodes.

5. A neural network radar processor as recited in claim 1, further comprising:

a. a direct digital synthesizer for generating a first oscillation signal having a first frequency of oscillation;

b. an intermediate frequency source for generating a second oscillation signal having a first frequency of oscillation;

c. a first mixer operatively connected to said direct digital synthesizer and to said intermediate frequency source for generating from said first and second oscillation signals a third oscillation signal having a third frequency of oscillation;

d. a direct reference oscillator for generating a fourth oscillation signal having a fourth frequency of oscillation

e. a second mixer operatively connected to said direct reference oscillator and to the output of said first mixer for generating from said third and fourth oscillation signals a fifth oscillation signal having a fifth frequency of oscillation;

f. a circulator operatively connected to the output of said second mixer;

g. at least one transmitting antenna for illuminating the object with a transmitted RF signal, whereby said at least one transmitting antenna is operatively connected to said circulator, whereby said circulator couples said fifth oscillation signal into said at least one transmitting antenna, said at least one transmitting antenna radiates said fifth oscillation signal;

h. at least one receiving antenna for receiving as a received RF signal a component of said transmitted RF signal reflected by the object, whereby said at least one receiving antenna is operatively connected to said circulator, said at least one receiving antenna receives said reflected component;

i. a third mixer operatively connected to said circulator and to said direct reference oscillator, whereby said circulator couples said received RF signal into said third mixer and said third mixer generates from said received RF signal and said fourth oscillation signal an intermediate RF signal;

j. a fourth mixer operatively connected to said direct digital synthesizer and to the output of said third mixer for generating from said first oscillation signal and said intermediate RF signal detected RF signal;

k. a quadrature phase shifter operatively connected to the output of said fourth mixer for shifting the phase of said detected RF signal by ninety degrees so as to create a phase-shifted detected RF signal;

l. a first analog-to-digital converter for sampling said detected RF signal, so as to generate said first time series;

m. a second analog-to-digital converter for sampling said phase-shifted detected RF signal, so as to generate said second time series; and

n. a signal processor operatively connected to said first and second analog-to-digital converters and to said direct digital synthesizer, whereby said signal processor controls the operating frequency of said direct digital synthesizer responsive to a ode from said signal processor, said code comprises a plurality of levels, each said level maintained for a period of time, said plurality of levels are substantially uniformly spaced, said period of time is sufficient for the propagation of said transmitted RF signal from said transmitting antenna to a target to said receiving antenna, and said code is repeated as a repetitive sequence.

6. A neural network radar processor as recited in claim 5, wherein said code comprises a first plurality of steps of increasing frequency followed by a second plurality of steps of decreasing frequency.

7. A neural network radar processor as recited in claim 6, wherein said code further comprises a dwell period between said first and second pluralities of steps.

8. A method of processing a radar signal with a neural network, comprising:

a. forming a time series by time sampling a radar return signal, wherein said time series comprises a plurality of samples;

b. forming a first plurality of first weighted sums, wherein each first weighted sum comprises a sum of first values weighted by a corresponding set of first weights, and each first value is a first function of a different sample of said time series;

c. forming a second plurality of second weighted sums, wherein each second weighted sum comprises a sum of second values weighted by a corresponding set of second weights, said first plurality of first weighted sums comprises a third plurality of mutually exclusive subsets of said first weighted sums, said second plurality of second weighted sums comprises a fourth plurality of mutually exclusive subsets of said second weighted sums, and each said second value of a second weighted sum of one of said fourth plurality of mutually exclusive subsets is a second function of a different first weighted sum of only one of said third plurality of mutually exclusive subsets of first weighted sums; and

d. adjusting said first and second sets of weights so that said third set of weighted sums of one of said fourth plurality of mutually exclusive subsets is substantially responsive only to targets in a first set of target ranges, and said third set of weighted sums of another of said fourth plurality of mutually exclusive subsets is substantially responsive only to targets in a second set of target ranges, whereby said first and second sets of target ranges are not overlapping.

9. A method of processing a radar signal with a neural network as recited in claim 8, wherein the operation of forming a time series comprises time sampling in-phase and quadrature-phase components of a down-converted radar return signal.

10. A method of processing a radar signal with a neural network as recited in claim 8, wherein said first function comprises at least one first parameter and said second function comprises at least one second parameter, further comprising the operation of adjusting said at least one first parameter and said at least one second parameter so that said third set of weighted sums of one of said fourth plurality of mutually exclusive subsets is substantially responsive only to targets in a first set of target ranges, and said third set of weighted sums of another of said fourth plurality of mutually exclusive subsets is substantially responsive only to targets in a second set of target ranges, whereby said first and second sets of target ranges are not overlapping.

11. A method of training a neural network in a neural network radar processor, comprising:

a. forming at least one first time series representing a radar return signal from a target space, wherein each said component comprises an associated waveform and said at least one first time series comprises a plurality of samples;

b. applying said at least one first time series to an input layer of a neural network, wherein said input layer comprises a first plurality of first nodes, one of said plurality of samples is applied to each of said first plurality of first nodes, a different of said plurality of samples is applied to different said first nodes, said neural network comprises a second plurality of nodes, a third plurality of corresponding weight values operative at said second plurality of nodes, and a fourth plurality of output nodes; and

c. adjusting said third plurality of corresponding weight values so that an output signal at each of said fourth plurality of output nodes is representative of said target space, wherein said at least one first time series is selected from a time series representative of a target space, a time series with a leakage signal and a time series with a DC bias;

said target space is selected from a target space with a variation of target reflection size, a target space with target fluctuation, a target space with a variation of target position relative to said radar, a target space with a variation in target speed relative to said radar, a target space with a variation in a target viewing angle, a target space with a variation in target shape, a target space with a variation in target size, a target space with a variation in antenna beam pointing angle, a target space with background clutter, and a target space with glint; and

for said at least one first time series selected from a time series with a leakage signal but no target, a time series with a DC bias but no target, and a time series representing a target space with background clutter, said third plurality of corresponding weight values are adjusted so that said output signal at each of said fourth plurality of output nodes tend towards a nullity.

12. A method of training a neural network in a neural network radar processor as recited in claim 11, wherein said operation of forming at least one first time series comprises forming at least one first time series of in-phase and quadrature-phase components representing a down-converted radar return signal from a target space.

13. A method of training a neural network in a neural network radar processor as recited in claim 11, wherein said operation of adjusting said third plurality of corresponding weight values comprises a back propagation process.

14. A method of training a neural network in a neural network radar processor as recited in claim 11, wherein the operation of adjusting said third plurality of corresponding weight values comprises adjusting said third plurality of corresponding weight values so that for said at least one first time series selected from a time series with a leakage signal but no target, a time series with a DC bias but no target, and a time series representing a target space with background clutter, said output signal at each of said fourth plurality of output nodes tend towards a nullity.

15. A method of training a neural network in a neural network radar processor as recited in claim 11, wherein said neural network further comprises at least one function with at least one parameter, further comprising the operation of adjusting said at least one parameter so that for said at least one first time series selected from a time series with a leakage signal but no target, a time series with a DC bias but no target, and a time series representing a target space with background clutter, said output signal at each of said fourth plurality of output nodes tend towards a nullity.

16. A method of training a neural network in a neural network radar processor as recited in claim 11, wherein for said at least one time series representative of a target space having a target at a range, said operation of adjusting said third plurality of corresponding weight values further comprises adjusting said third plurality of corresponding weight values so that said output signal of an output node of said fourth plurality of output nodes corresponding to said range of said target tends towards a value indicative of said target at said range, and said output signals for other output nodes of said fourth plurality of output nodes tend towards a nullity if no target is present at a range corresponding to said other output nodes.

17. A method of training a neural network in a neural network radar processor as recited in claim 11, wherein said neural network further comprises at least one function with at least one parameter, further comprising the operation of adjusting said at least one parameter so that for said at least one time series representative of a target space having a target at a range, said operation of adjusting said at least one parameter further comprises adjusting said at least one parameter so that said output signal of an output node of said fourth plurality of output nodes corresponding to said range of said target tends towards a value indicative of said target at said range, and said output signals for other output nodes of said fourth plurality of output nodes tend towards a nullity if no target is present at a range corresponding to said other output nodes.

18. A method of training a neural network in a neural network radar processor as recited in claim 11, wherein the operation of forming said at least one first time series comprises an operation selected from the operations of 1) replaying a recording of a radar return signal from a radar system and 2) forming said at least one first time series of a simulated radar return signal.