Classification systems and methods using convex hulls

Patent metadata

Field	Value
Publication number	US-9600765-B1
Application number	US-201615076040-A
Country	US
Kind code	B1
Filing date	Mar 21, 2016
Priority date	Sep 27, 2012
Publication date	Mar 21, 2017
Grant date	Mar 21, 2017

Abstract

A computerized classifier system that forms convex hulls containing all training experiences for each target class (e.g. threat/non-threat) is disclosed. The intersection of convex hulls for each pair of classes defines a region of ambiguity in feature space. Measurement of feature vector(s) outside an ambiguous region of feature space leads to a class decision while measurement of feature vector(s) within an ambiguous region of feature space defined by convex hulls causes a transition to a new feature space with new features. In particular embodiments, measured feature data includes estimated motion states and electrical lengths of a given object, and range, velocity and acceleration image data from second phase differences for debris mitigation.

First claim

What is claimed is:

1. A computer implemented method for determining a change in state of a target being monitored comprising the steps of:
measuring feature data associated with said target during a defined time duration;
processing using a computer processor, real time sensor data from said target using a neural network data structure that has defined therein convex hulls having regions of ambiguity corresponding to each of a set of target classes corresponding to target features, wherein the intersection of convex hulls for each pair of classes defines the region of ambiguity in feature space, wherein the neural network data structure uses training data to receive the real time sensor data and recognize signal characteristics which indicate the presence of information relating to a target class designation;
transitioning, according to the neural network data structure, to another feature space having different features, in response to the neural network data structure's processing of measurement data of feature vectors within an ambiguous region of feature space defined by convex hulls; and
determining using a computer processor, a next sensor measurement data for obtaining additional measured feature data of the target according to a current state of the graph data structure and the region of ambiguity; and
repeating said processing, transitioning and determining steps until the change in state of the target is classified.

2. The method of claim 1 , wherein the neural network data structure comprises:
a layer of input nodes;
a layer of output nodes; and
and one or more layers of processing nodes, wherein a training algorithm for the neural network data structure comprises processing the real time sensor data to minimize an error in a pre-output layer of the neural network data structure that a detected target belongs to an assigned target class designation.

3. The method of claim 1 , wherein the neural network data structure is configured to process the real time sensor data to utilize motion-invariant generalized coordinates to assign a target class designation to the target being monitored.

4. The method of claim 3 , wherein the neural network data structure is further configured to process the real time sensor data and calculate a phase term that is representative of a remote scatterer and measurement data relating to a data cube within the measuring range of sensors obtaining the real time sensor data.

5. The method of claim 4 , further comprising calculating by the neural network data structure, a column vector by combining measured data relating to the data cube with the phase term representative of the remote scatterer.

6. The method of claim 5 , wherein the phase term is calculated according to:

ϕ
⁡

(

q
,
t
,
f
,
e

)

=

-

2
⁢
π
⁢

⁢
f

c

⁡

[



T
⁡

(
t
)

-

x
(
t


⁢
q

)

⁢


+


⁢

R
⁡

(
t
)

-

x
(

t
⁢


q
)



]

where t represents a pulse time of the sensor,
f represents frequency,
e represents a receiving element in a coherent phased sensor array,
q is a parameter which describes kinematic states of scatterers, and
T and R are vectors representing motion of the transmitter and receiver sensor, respectively.

7. The method of claim 6 , wherein the column vector is calculated according to:
| q,z = z ( t,f,e ) e −φ(q,t,f,e)
where z(t,f,e) is the measured data for a data cube for a given pulse time, frequency and sensor element.

8. A computer implemented method of mitigating debris and simultaneously estimating range, velocity and acceleration of at least one target using only linear discrete Fourier transform (DFT) processing comprising the steps of:
transmitting from a radar, radar signals into an area to be monitored;
receiving radar signals at a receiver, the radar signals reflected off of at least one target in the area to be monitored and including information relating to the at least one target;
modeling each target as a point having a constant velocity and a constant acceleration for a given coherent processing interval (CPI);
determining a two-way phase to each target for each frequency in a wideband frequency range of the radar;
calculating a phase difference for receive in-phase (I) and quadrature (Q) data for each frequency in the wideband frequency range over time to form a modified sequence of I and Q data;
modeling the phases of the modified sequence of I and Q data to produce a phase factor that is linear over time;
calculating a velocity-acceleration image using only a 2-dimensional DFT based on the modeled phases of the modified sequence of I and Q data.

9. The computer implemented method of claim 8 , further comprising the steps of:
based on the calculated velocity-acceleration image, detecting targets exhibiting ballistic acceleration in the velocity-acceleration image;
rejecting the detected targets exhibiting ballistic acceleration from the velocity-acceleration image;
calculating range, velocity and acceleration for each non-ballistic target according to non-linear equations, wherein velocity and

Assignees

• Lockheed Corp

Inventors

• Pedersen Richard N
• Friesel Mark A
• Mountcastle Paul D

Classifications

• G06N3/088 Primary

  Non-supervised learning, e.g. competitive learning · CPC title

• G06V10/765

  using rules for classification or partitioning the feature space · CPC title

• G06V10/82 Primary

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/24155

  Bayesian classification · CPC title