Motion detection based on machine learning of wireless signal properties

• US 10,108,903 B1
• Filed: 12/08/2017
• Issued: 10/23/2018
• Est. Priority Date: 12/08/2017
• Status: Active Grant

• Current Assignee: Cognitive Systems Corp.
• Original Assignee: Cognitive Systems Corp.
• Inventors: Piao, Yunfeng, Olekas, Christopher Vytautas, Zakharov, Mikhail Alexand
• Primary Examiner(s): Misir, Dave
• Application Number:
  US15/836,606
• Time in Patent Office: 319 Days
• Field of Search: 706 20
• CPC Class Codes: G01S 13/003 Bistatic radar systems; Multistatic radar systems
• CPC Class Codes: G01S 13/04 Systems determining presence of a target based on relative movement of target G01S13/56
• CPC Class Codes: G01S 13/765 with exchange of information between interrogator and responder
• CPC Class Codes: G01S 13/886 for alarm systems alarms with electrical actuation G08B13/22
• CPC Class Codes: G01S 7/417 involving the use of neural networks
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/0464 Convolutional networks [CNN, ConvNet]
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G08B 13/2494 by interference with electro-magnetic field distribution combined with other electrical sensor means, e.g. microwave detectors combined with other sensor means

Abstract

In a general aspect, motion in a space can be detected based on machine learning of wireless signal properties. In some aspects, sets of tagged neural network input data are obtained at a neural network training system. Each set of tagged neural network input data is based on a statistical analysis of a series of wireless signals transmitted through a space over a respective time period, and each set of the tagged neural network input data includes a tag indicating whether motion occurred in the space over the respective time period. The sets of tagged neural network input data are processed by the neural network training system to parameterize nodes of a neural network system. Parameterizing the nodes configures the neural network system to detect motion based on untagged neural network input data.

Claims

1. A motion detection method comprising:

obtaining, at a neural network training system, multiple sets of tagged neural network input data, each set of tagged neural network input data based on a statistical analysis of a series of wireless signals transmitted through a space over a respective time period, each set of the tagged neural network input data comprising a tag indicating whether motion occurred in the space over the respective time period,

wherein the tagged neural network input data comprises histogram data, and the statistical analysis comprises;

obtaining a frequency-domain representation of the wireless signals,

computing statistical parameter values based on the frequency-domain representation of the wireless signals,

populating the statistical parameter values into an initial matrix, and

generating the histogram data based on the initial matrix, the histogram data comprising a set of bins and a quantity for each bin, each bin corresponding to a respective range for each of the statistical parameters; and

by operation of the neural network training system, processing the sets of tagged neural network input data to parameterize nodes of a neural network system;

detecting motion, using the neural network system comprising the parameterized nodes, based on untagged neural network input data; and

activating a security system or a physical device associated with the space where the motion was detected.

2. The method of claim 1, wherein the histogram data comprises an image of the statistical parameters plotted against one or more axes.

3. The method of claim 1, wherein the statistical parameters comprise a mean and a standard deviation.

4. The method of claim 1, wherein the tag for each set of the tagged neural network input data indicates a category of motion, and parameterizing the nodes configures the neural network system to detect a category of motion based on the untagged neural network input data.

5. The method of claim 1, comprising:

obtaining, at the neural network training system, additional sets of tagged neural network input data, each additional set of tagged neural network input data comprising a tag indicating whether interference was present in the space over the respective time period; and

by operation of the neural network training system, processing the additional sets of tagged neural network input data to parameterize nodes of the neural network system, wherein parameterizing the nodes configures the neural network system to detect interference based on the untagged neural network input data.

6. The method of claim 1, wherein the neural network system comprises a convolutional neural network.

7. A neural network training system comprising:

a data processing apparatus; and

memory comprising instructions that are operable when executed by the data processing apparatus to perform operations comprising;

obtaining multiple sets of tagged neural network input data, each set of tagged neural network input data based on a statistical analysis of a series of wireless signals transmitted through a space over a respective time period, each set of the tagged neural network input data comprising a tag indicating whether motion occurred in the space over the respective time period,

wherein the tagged neural network input data comprises histogram data, and the statistical analysis comprises;

obtaining a frequency-domain representation of the wireless signals;

computing the statistical parameter values based on the frequency-domain representation of the wireless signals;

populating the statistical parameter values into an initial matrix; and

generating the histogram data based on the initial matrix, the histogram data comprising a set of bins and a quantity for each bin, each bin corresponding to a respective range for each of the statistical parameters; and

processing the sets of tagged neural network input data to parameterize nodes of a neural network system;

detecting motion, using the neural network system comprising the parameterized nodes, based on untagged neural network input data; and

activating a security system or a physical device associated with the space where the motion was detected.

8. The neural network training system of claim 7, wherein the histogram data comprises an image of the statistical parameters plotted against one or more axes.

9. The neural network training system of claim 7, wherein the statistical parameters comprise a mean and a standard deviation.

10. The neural network training system of claim 7, wherein the tag for each set of the tagged neural network input data indicates a category of motion, and parameterizing the nodes configures the neural network system to detect a category of motion based on the untagged neural network input data.

11. The neural network training system of claim 7, wherein the operations comprise:

obtaining additional sets of tagged neural network input data, each additional set of tagged neural network input data comprising a tag indicating whether interference was present in the space over the respective time period; and

processing the additional sets of tagged neural network input data to parameterize nodes of the neural network system, wherein parameterizing the nodes configures the neural network system to detect interference based on the untagged neural network input data.

12. The neural network training system of claim 7, wherein the neural network system comprises a convolutional neural network.

13. A motion detection method, comprising:

obtaining, at a neural network system, multiple sets of neural network input data, each set of neural network input data based on a statistical analysis of a series of wireless signals transmitted through a space over a respective time period,

wherein the neural network input data comprises histogram data, and the statistical analysis comprises;

obtaining a frequency-domain representation of the wireless signals;

computing statistical parameter values based on the frequency-domain representation of the wireless signals;

populating the statistical parameter values into an initial matrix; and

generating the histogram data based on the initial matrix, and

by operation of the neural network system, processing the sets of neural network input data through a plurality of programmed neural network nodes;

determining whether motion occurred in the space during the respective time period; and

responsive to determining motion occurred in the space, activating a security system or a physical device associated with the space where the motion was detected.

14. The method of claim 13, wherein the statistical parameters comprise a mean and a standard deviation.

15. The method of claim 13, wherein determining whether motion occurred in the space comprises generating an indication of motion by an object in the space, a category of motion that occurred in the space, interference present in the space, or an absence of motion in the space.

16. The method of claim 13, wherein the neural network system comprises a convolutional neural network and a fully-connected neural network.

17. The method of claim 16, wherein the convolutional neural network comprises:

multiple convolutional layers;

a max-pooling layer after at least one of the convolutional layers;

a flattening layer after the max-pooling layer; and

multiple dense layers after the flattening layer.

18. The method of claim 16, wherein processing the neural network input data through a plurality of programmed neural network nodes comprises:

detecting, by the convolutional neural network, that motion occurred for a subset of the neural network input data;

computing mean square error values for the subset of the neural network input data; and

providing the mean square error values as input to the fully-connected neural network to confirm whether motion did occur.

19. A neural network system comprising:

a data processing apparatus; and

memory comprising instructions that are operable when executed by the data processing apparatus to perform operations comprising;

obtaining multiple sets of neural network input data, each set of neural network input data based on a statistical analysis of a series of wireless signals transmitted through a space over a respective time period,

wherein the neural network input data comprises histogram data, and the statistical analysis comprises;

obtaining a frequency-domain representation of the wireless signals;

computing statistical parameter values based on the frequency-domain representation of the wireless signals;

populating the statistical parameter values into an initial matrix; and

generating the histogram data based on the initial matrix; and

processing the sets of neural network input data through a plurality of programmed neural network nodes;

determining whether motion occurred in the space during the respective time period; and

responsive to determining motion occurred in the space, activating a security system or a physical device associated with the space where the motion was detected.

20. The neural network system of claim 19, wherein the statistical parameters comprise a mean and a standard deviation.

21. The neural network system of claim 19, wherein determining whether motion occurred in the space comprises generating an indication of motion by an object in the space, a category of motion that occurred in the space, interference present in the space, or an absence of motion in the space.

22. The neural network system of claim 19, wherein the neural network system comprises a convolutional neural network and a fully-connected neural network.

23. The neural network system of claim 22, wherein the convolutional neural network comprises:

multiple convolutional layers;

a max-pooling layer after at least one of the convolutional layers;

a flattening layer after the max-pooling layer; and

multiple dense layers after the flattening layer.

24. The neural network system of claim 22, wherein processing the neural network input data through a plurality of programmed neural network nodes comprises:

detecting, by the convolutional neural network, that motion occurred for a subset of the neural network input data; and

computing mean square error values for the subset of the neural network input data; and

providing the mean square error values as input to the fully-connected neural network to confirm whether motion did occur.