Method for tracking multiple target objects, device, and computer program for implementing the tracking of multiple target objects for the case of moving objects

Patent metadata

Field	Value
Publication number	US-11253997-B2
Application number	US-201916260703-A
Country	US
Kind code	B2
Filing date	Jan 29, 2019
Priority date	Feb 1, 2018
Publication date	Feb 22, 2022
Grant date	Feb 22, 2022

Abstract

A method for tracking multiple target objects, moving objects to be tracked being projected onto a grid map having grid cells, the method including the following tasks to be executed in each time: computing the velocity distribution for the next time step with the aid of a transition velocity distribution, which indicates how the objects associated with a grid cell in question move from one time step to the next, based on the preceding velocity distribution; for each grid cell, calculating a transitional probability information item, which indicates, for objects in each grid cell, probabilities of the objects in question reaching possible, further grid cells, as a function of the velocity distribution; calculating an occupancy probability for each grid cell for a subsequent time, based on the transitional probability information item; operating a system as a function of the occupancy probabilities for the grid cells.

First claim

What is claimed is:

1. A method for tracking multiple target objects, moving objects to be tracked being projected onto a grid map having grid cells, the method comprising:
computing a velocity distribution for a next time step with a transitional velocity distribution, which indicates how the objects associated with a grid cell move from one time step to the next time step, based on the preceding velocity distribution;
calculating, for each of the grid cells, a transitional probability information item, which indicates, for objects in each of the grid cells, probabilities of the objects reaching possible, further grid cells, as a function of the velocity distribution;
calculating an occupancy probability for each of the grid cells for a subsequent time step, based on transitional probability information item; and
operating a system as a function of the occupancy probabilities for the grid cells wherein:
the transitional velocity distribution for each of the grid cells is calculated as a function of a supplied plan, with a trained neural network and with a multivariate Gaussian model,
the transitional velocity distribution is ascertained from an acceleration distribution, with a neighboring transitional velocity distribution obtained with the aid of the neural network, using the multivariate Gaussian model, and
the transitional velocity distribution for a particular grid cell is ascertained, in that:
an acceleration distribution is ascertained from the neighboring transitional velocity distribution by subtraction;
the acceleration distribution is expanded to a range of grid cells with the aid of the multivariate Gaussian model, and rendered discrete for grid cells situated about the particular grid cell; and
the velocity distribution for the next time step from the expanded acceleration distribution in the grid cells situated about the particular grid cell is ascertained;

a robot is controlled as a function of the occupancy probabilities for the grid cells.

2. The method of claim 1 , wherein a velocity distribution is initially provided for each of the grid cells.

3. The method of claim 1 , wherein the neural network is provided, in that:
measurement data regarding occupancies of grid cells over a number of time steps are provided in one or more training environments, which are described by one or more training plans;
for each of the grid cells, frequencies for each combination of occupancies of adjacent grid cells are determined, the frequencies each indicating how often the condition is satisfied, that a particular, adjacent grid cell is occupied in the preceding time step, that the particular grid cell is occupied in the current time step, and that a particular, adjacent grid cell is occupied in a subsequent time step;
probabilities are ascertained from the frequencies for each of the grid cells, so as to obtain a neighboring transitional velocity distribution for training; and
the neural network, in particular, of a convolutional neural network, is trained, using the neighboring transitional velocity distribution for training and the one or more training plans.

4. The method of claim 3 , wherein the neighboring transitional velocity distribution for each of the grid cells of the grid map of the defined local environment is modeled as a function of the specified plan and the trained neural network.

5. The method of claim 1 , wherein the measurement indicates an occupancy and/or velocity of one or more of the grid cells in a current time step, and the occupancy probability is corrected based on the measurement.

6. An apparatus for executing a method of tracking multiple target objects, moving objects to be tracked being projected onto a grid map having grid cells, comprising:
a device configured to perform the following:
calculating a velocity distribution for the next time step with a transition velocity distribution, which indicates how the objects associated with a particular grid cell move from one time step to the next, based on the preceding velocity distribution;
calculating, for each of the grid cells, a transitional probability information item, which indicates, for objects in each of the grid cells, probabilities of the objects reaching possible, further grid cells, as a function of the velocity distribution;
calculating an occupancy probability for each of the grid cells for a subsequent time step, based on the transitional probability information item; and

operating a system as a function of the occupancy probabilities for the grid cells wherein:
the transitional velocity distribution for each of the grid cells is calculated as a function of a supplied plan, with a trained neural network and with a multivariate Gaussian model, and
the transitional velocity distribution is ascertained from an acceleration distribution, with a neighboring transitional velocity distribution obtained with the aid of the neural network, using the multivariate Gaussian model, and
the transitional velocity distribution for a particular grid cell is ascertained, in that:
an acceleration distribution is ascertained from the neighboring transitional velocity distribution by subtraction;
the acceleration distribution is expanded to a range of grid cells with the aid of the multivariate Gaussian model, and rendered discrete for grid cells situated about the particular grid cell; and
the velocity distribution for the next time step from the expanded acceleration distribution in the grid cells situated about the particular grid cell is ascertained;

a robot is controlled as a function of the occupancy probabilities for the grid cells.

7. A non-transitory computer readable medium having a computer program, which is executable by a processor, comprising:
a program code arrangement having program code for tracking multiple target objects, moving objects to be tracked being projected onto a grid map having grid cells, tby performing the following:
computing a velocity distribution for a next time step with a transitional velocity distribution, which indicates how the objects associated with a grid cell move from one time step to the next time step, based on the preceding velocity distribution;
calculating, for each of the grid cells, a transitional probability information item, which indicates, for objects in each of the grid cells, probabilities of the objects reaching possible, further grid cells, as a function of the velocity distribution;
calculating an occupancy probability for each of the grid cells for a subsequent time step, based on transitional probability information item; and

operating a system as a function of the occupancy probabilities for the grid cells wherein:
the transitional velocity distribution for each of the grid cells is calculated as a function of a supplied plan, with a trained neural network and with a multivariate Gaussian model, and
the transitional velocity distribution is ascertained from an acceleration distribution, with a neighboring transitional velocity distribution obtained with the aid of the neural network, using the multivariate Gaussian model, and
the transitional velocity distribution for a particular grid cell is ascertained, in that:
an acceleration distribution is ascertained from the neighboring transitional velocity distribution by subtraction;
the acceleration distribution is expanded to a range of grid cells with the aid of the multivariate Gaussian model, and rendered discrete for grid cells situated about the particular grid cell; and
the velocity distribution for the next time step from the expanded acceleration distribution in the grid cells situated about the particular grid cell is ascertained;

a robot is controlled as a function of the occupancy probabilities for the grid cells.

Assignees

• Bosch Gmbh Robert

Inventors

• Spies Markus
• Doellinger Johannes Maximilian
• Fu Liangcheng

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G05B13/027

  using neural networks only · CPC title

• B25J9/1676

  Avoiding collision or forbidden zones · CPC title

• B25J9/161 Primary

  Hardware, e.g. neural networks, fuzzy logic, interfaces, processor · CPC title