More reliable classification of radar data from dynamic settings

Patent metadata

Field	Value
Publication number	US-11835620-B2
Application number	US-202017119608-A
Country	US
Kind code	B2
Filing date	Dec 11, 2020
Priority date	Dec 18, 2019
Publication date	Dec 5, 2023
Grant date	Dec 5, 2023

Abstract

A method for classifying radar data, which have been obtained by registering radar radiation emitted from a transmitter and reflected by at least one object using at least one detector. The method includes: providing radar data, which include observations of a setting recorded at different points in time; ascertaining at least one portion of the radar data, which is rotated and/or scaled in at least one of the observations as compared to at least one other of the observations; ascertaining a fixed point of the rotation and/or scaling; transforming at least one two-dimensional representation of at least one part of the observations into logarithmic polar coordinates using the ascertained fixed point as the origin; and mapping the at least one transformed two-dimensional representation onto at least one class of a predefined classification via at least one classifier, which encompasses a neural network that includes at least one convolution layer.

First claim

What is claimed is:

1. A method for classifying radar data, which have been obtained by registering radar radiation emitted from a transmitter and reflected by at least one object using at least one detector, the method comprising the following steps:
providing radar data, which include observations of a setting recorded at different points in time;
ascertaining at least one portion of the radar data, which is rotated and/or scaled in at least one of the observations as compared to at least one other of the observations;
ascertaining a fixed point of the rotation and/or scaling;
transforming at least one two-dimensional representation of at least one part of the observations into logarithmic polar coordinates using the ascertained fixed point as the origin; and
mapping the at least one transformed two-dimensional representation onto at least one class of a predefined classification via at least one classifier, the at least one classifier encompassing a neural network that includes at least one convolution layer.

2. The method as recited in claim 1 , wherein two-dimensional or three-dimensional spatial distributions of at least two measured variables resulting from the reflected radar radiation are combined in a multi-dimensional tensor, two or three dimensions of the tensor representing a spatial dimensions of the distributions and a further dimension of the tensor representing the number of available measured variables.

3. The method as recited in claim 2 , wherein the measured variables include:
an intensity of the reflected radar radiation, and/or
a radar cross section of objects in the setting, and/or
a velocity component of the objects in a propagation direction of the radar radiation.

4. The method as recited in claim 1 , wherein the ascertainment of the fixed point includes assessing each of multiple candidate fixed points using a quality function in terms of how well it conforms to the observations, and selecting the fixed point based on the assessments.

5. The method as recited in claim 4 , wherein a parameterized approach including free parameters is established for the candidate fixed points, and the free parameters are optimized with an aim of the quality function assuming an extremum.

6. The method as recited in claim 4 , wherein multiple observations of the observations are fed to a trained classifier and/or regressor, which maps the multiple observations onto the fixed point.

7. The method as recited in claim 1 , wherein the radar data are selected, which have been detected using a detector mounted at or in a vehicle, and wherein classes of the predefined classification represent traffic signs, and/or other road users, and/or traffic lane boundaries, and/or obstacles, and/or other traffic-relevant objects.

8. The method as recited in claim 7 , further comprising:
generating an activation signal as a function of the at least one class provided by the classifier, and activing the vehicle using the activation signal.

9. The method as recited in claim 1 , wherein the radar data are selected, which have been detected by irradiating an area to be monitored with radar radiation from a stationary transmitter and by measuring the reflected radiation using at least one stationary detector, and wherein classes of the predefined classification representing persons, and/or animals, and/or vehicles, and/or tools, and/or other objects relevant for safety of an area to be monitored.

10. The method as recited in claim 9 , further comprising:
generating an activation signal as a function of the at least one class provided by the classifier, and activating a system for monitoring the area using the activation signal.

11. A method for training a classifier, the classifier encompasses a neural network including at least one convolution layer, the method comprising the following steps:
providing learning radar data with learning observations;
providing setpoint classes of a predefined classification, which are assigned to each of the learning observations;
transforming at least one two-dimensional representation each of at least one part of each of the learning observations into logarithmic polar coordinates including a fixed point as the origin, the fixed point being selected as a point around which observations in an intended area of the classifier probably appear rotated and/or scaled as compared to the learning observations;
mapping the transformed two-dimensional representations onto one or multiple classes of the predefined classification via the classifier; and
optimizing parameters which characterize a behavior of the classifier, with an aim that the classes provided by the classifier to the transformed representations of the learning observations coincide with the setpoint classes according to a predefined cost function.

12. A non-transitory machine-readable data medium on which is stored a computer program for classifying radar data, which have been obtained by registering radar radiation emitted from a transmitter and reflected by at least one object using at least one detector, the computer program, when executed by one or more computers, causing the one or more computers to perform the following steps:
providing radar data, which include observations of a setting recorded at different points in time;
ascertaining at least one portion of the radar data, which is rotated and/or scaled in at least one of the observations as compared to at least one other of the observations;
ascertaining a fixed point of the rotation and/or scaling;
transforming at least one two-dimensional representation of at least one part of the observations into logarithmic polar coordinates using the ascertained fixed point as the origin; and
mapping the at least one transformed two-dimensional representation onto at least one class of a predefined classification via at least one classifier, the at least one classifier encompassing a neural network that includes at least one convolution layer.

13. A computer configured to classify radar data, which have been obtained by registering radar radiation emitted from a transmitter and reflected by at least one object using at least one detector, the computer configured to:
provide radar data, which include observations of a setting recorded at different points in time;
ascertain at least one portion of the radar data, which is rotated and/or scaled in at least one of the observations as compared to at least one other of the observations;
ascertain a fixed point of the rotation and/or scaling;
transform at least one two-dimensional representation of at least one part of the observations into logarithmic polar coordinates using the ascertained fixed point as the origin; and
map the at least one transformed two-dimensional representation onto at least one class of a predefined classification via at least one classifier, the at least one classifier encompassing a neural network that includes at least one convolution layer.

Assignees

• Bosch Gmbh Robert

Inventors

• Natroshvili Koba

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G01S13/91 Primary

  for traffic control (G01S13/93 takes precedence) · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G09B9/54

  Simulation of radar (G09B9/40 takes precedence) · CPC title