Granted Patent B2

US 11,308,335 · App. 16/876,535 · Granted Apr 19, 2022

Inventors: Timothy Sulzer (Jenkintown, PA); Michael Lahiff (Jenkintown, PA); Marcus Day (Jenkintown, PA)

Assignee: ZeroEyes, Inc.

G06K9/00771G06K9/3233G06K9/6255G06K9/6257G06K9/6262H04N7/183

Abstract

Claims (65)

1. A method for training an object detection device, the method comprising the steps of:

recording a first video stream from a security camera;

selecting a set of frames from the first video stream;

selecting a subset of frames from the set of frames and inserting a bounding box in each frame of the subset of frames;

creating a list of attributes associated with said each frame of the subset of frames;

creating a database comprising said each frame of the subset of frames and the associated list of attributes, wherein the database is searchable by at least one attribute of the list of attributes;

training a detection model using the database, wherein the training includes varying a first parameter of the detection model;

analyzing a second video stream using the detection model to determine a number of true positive events and a number of false positive events, wherein the second video stream is different from the first video stream;

creating a dataset of the analysis of the second video stream;

filtering the dataset using a second parameter;

determining a ratio of true positive events to false positive events (“TP/FP”);

determining a ratio of false positive events to true positive events (“FP/TP”);

converting the FP/TP to a percentage (“% FP/TP); and

evaluating a performance of the detection model based at least in part on a ratio of % FP/TP to TP/FP.

2. The method of claim 1 , wherein the first video stream comprises an image of a person carrying an object of interest.

3. The method of claim 1 , wherein the object of interest is a weapon.

4. The method of claim 1 , wherein at least one frame of the subset of frames includes an object of interest.

5. The method of claim 4 , wherein the bounding box circumscribes at least a part of the object of interest.

6. The method of claim 1 , wherein the bounding box is a bounding polygon.

7. The method of claim 1 , wherein at least one of the attributes is associated with an object of interest.

8. The method of claim 1 , wherein the attributes are selected from the group consisting of: color, lighting level, clarity, security camera characteristics, an object of interest type, an orientation of the object of interest, contrast, and combinations thereof.

9. The method of claim 1 further comprising the step of adjusting a location of the bounding box in one of the frames of the subset of frames from a first location to a second location.

10. The method of claim 9 , wherein the first location and the second location are separated by a predetermined number of pixels.

11. The method of claim 1 , further comprising the step of adjusting a size of the bounding box in one of the frames of the subset of frames from a first size to a second size.

12. The method of claim 11 , wherein the first size of the bounding box comprises a first number of pixels and the second size of the bounding box comprises a second number of pixels different from the first number of pixels.

13. The method of claim 1 , wherein the first parameter is selected from the group consisting of: a number of labels, a number of images, a number of frames, a number of iterations, a max iteration value, a test iteration value, a test interval, a momentum value, a ratio value, a learning rate, a weight decay, and combinations thereof.

14. The method of claim 1 , wherein the second video stream comprises a live video stream or a filtered video stream.

15. The method of claim 1 , wherein the second parameter is selected from the group consisting of: confidence score, event duration, pixel area size, object speed, minimum range of object movement, average object size, and average pixel speed, and combinations thereof.

16. The method of claim 1 , further comprising the step of maximizing a difference between the number of true positive events and the number of false positive events, wherein the number of true positive events is greater than the number of false positive events.

17. The method of claim 1 , wherein the evaluation of the performance of the detection model comprises determining performance metrics.

18. The method of claim 17 , wherein the performance metrics are selected from the group consisting of: number of true positive events, number of false positive events, a score value, an average score value, a label performance value, a score by distance value, and combinations thereof.

19. The method of claim 1 , further comprising the step of comparing the evaluated performance of the detection model with a performance of a second detection model.

20. The method of claim 1 , further comprising the step of adjusting a location of the insertion of the bounding box in a frame of the subset of frames based on the evaluated performance of the detection model.

21. The method of claim 1 , further comprising the step of changing an attribute in the list of attributes based on the evaluated performance of the detection model.

22. The method of claim 1 , further comprising the step of adding or deleting an attribute from the list of attributes based on the evaluated performance of the detection model.

23. A system for training an object detection device, comprising:

a recording device for recording a first video stream from a security camera;

a processor operatively connected to said recording device, said processor programmed to select a set of frames from the first video stream;

said processor further programmed to select a subset of frames from the set of frames and inserting a bounding box in each frame of the subset of frames;

said processor further programmed to create a list of attributes associated with said each frame of the subset of frames;

a database operatively connected to said processor, wherein said processor populates said database such that said database comprises said each frame of the subset of frames and the associated list of attributes, wherein said database is searchable by at least one attribute of the list of attributes;

said processor further programmed to train a detection model using the database, wherein the training includes varying a first parameter of the detection model;

said processor further programmed to analyze a second video stream using the detection model to determine a number of true positive events and a number of false positive events, wherein the second video stream is different from the first video stream;

said processor further programmed to create a dataset of the analysis of the second video stream;

said processor further programmed to filter the dataset using a second parameter;

said processor further programmed to evaluate a performance of the detection model, at least in part, by:

determining a ratio of true positive events to false positive events (“TP/FP”),

determining a ratio of false positive events to true positive events (“FP/TP”),

converting the FP/TP to a percentage (“% FP/TP), and

determining a ratio of % FP/TP to TP/FP.

24. A non-transitory, machine-readable medium having stored thereon a plurality of

executable instructions, the plurality of executable instructions comprising instructions to:

record a first video stream from a security camera;

select a set of frames from the first video stream;

select a subset of frames from the set of frames and inserting a bounding box in each frame of the subset of frames;

create a list of attributes associated with said each frame of the subset of frames;

create a database comprising said each frame of the subset of frames and the associated list of attributes, wherein the database is searchable by at least one attribute of the list of attributes;

train a detection model using the database, wherein the training includes varying a first parameter of the detection model;

analyze a second video stream using the detection model to determine a number of true positive events and a number of false positive events, wherein the second video stream is different from the first video stream;

create a dataset of the analysis of the second video stream;

filter the dataset using a second parameter

determine a ratio of true positive events to false positive events (“TP/FP”);

determine a ratio of false positive events to true positive events (“FP/TP”);

convert the FP/TP to a percentage (“% FP/TP); and

evaluate a performance of the detection model based at least in part on a ratio of % FP/TP to TP/FP.

Assignments (9)

Continuity (2)

Cited By (2)

US 12,511,891 US 12,613,520