Training a generator unit and a discriminator unit for collision-aware trajectory prediction

Abstract

A system trains a generator unit and a discriminator unit simultaneously. The generator unit is configured to determine a future trajectory of at least one other road user in the environment of a vehicle considering an observed trajectory of the at least one other road user. The discriminator unit is configured to determine whether the determined future trajectory of the other road user is an actual future trajectory of the other road user. The system is configured to train the generator unit and the discriminator unit simultaneously with gradient descent.

Claims

1. A system, comprising: an artificial neural network comprising a generator unit, a discriminator unit, and an oracle unit, wherein said generator unit, said discriminator unit, and said oracle unit are executed by a computer, wherein said generator unit is configured to determine a future trajectory of at least one other road user in an environment of a vehicle considering an observed trajectory of the at least one other road user, wherein said discriminator unit is configured to determine whether the determined future trajectory of the at least one other road user is an actual future trajectory of the at least one other road user, and wherein said computer is configured to train said generator unit and said discriminator unit simultaneously with gradient descent, wherein said oracle unit is configured to determine a reward for the determined future trajectory of the at least one other road user considering whether the determined future trajectory of the other road user is collision-free, and wherein said computer is configured to train said generator unit considering the reward determined by the oracle unit.

2. The system according to claim 1, wherein the other road user is a vulnerable road user.

3. The system according to claim 1, wherein the generator unit is configured to determine the future trajectory of the at least one other road user considering at least one static object in the environment of the other road user.

4. The system according to claim 3, wherein the generator unit is configured to determine the future trajectory of the other road user considering the relative location of the at least one static object.

5. The system according to claim 3, wherein the generator unit is configured to determine the future trajectory of the other road user considering at least one dynamic object in the environment of the other road user.

6. The system according to claim 1, wherein the generator unit comprises an encoder unit, with said encoder unit configured to map an observed trajectory of the other road user to a common embedding space.

7. The system according to claim 6, wherein the encoder unit comprises a long short-term memory unit.

8. The system according to claim 6, wherein the generator unit comprises a decoder unit, with said decoder unit configured to determine the future trajectory of the other road user considering the common embedding space.

9. The system according to claim 8, wherein the decoder unit comprises a long short-term memory unit.

10. A generator unit trained by the system according to claim 1.

11. A computer implemented method for training a generator unit and a discriminator unit of an artificial neural network, wherein said generator unit is configured to determine a future trajectory of at least one other road user in the environment of a vehicle user considering an observed trajectory of the at least one other road user, said discriminator unit is configured to determine whether the determined future trajectory of the other road user is an actual future trajectory of the other road user, said artificial neural network includes an oracle unit that is configured to determine a reward for the determined future trajectory of the at least one other road user considering whether the determined future trajectory of the other road user is collision-free, and said generator unit, said discriminator unit, and said oracle unit are executed by a computer, the method comprising the step of: training, by the computer, said generator unit and said discriminator unit, wherein said training is carried out simultaneously with gradient descent, and said training considers the reward determined by the oracle unit.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 shows an exemplified structure of the system for training the artificial intelligence unit.

(2) FIG. 2 shows an exemplified grid in polar angle space.

DETAILED DESCRIPTION OF THE DRAWINGS

(3) FIG. 1 shows a system for training a generator unit GU and a discriminator unit DU simultaneously. Said generator unit GU is configured to determine a future trajectory FT of at least one other road in the environment of a vehicle considering an observed trajectory OT of the at least one other road user. The other road user is in particular a vulnerable road user, e.g. a pedestrian, a cyclist or an animal.

(4) The generator unit GU is in particular configured to determine the future trajectory FT of the at least one other road user considering at least one static object SO in the environment of the other road user, particular considering the relative location of the at least one static object SO.

(5) Moreover, the generator GU unit is in particular configured to determine the future trajectory FT of the other road user considering at least one dynamic object DO in the environment of the other road user.

(6) The generator unit GU comprises in particular an encoder unit EU, with said encoder unit EU configured to map an observed trajectory OT of the other road user to a common embedding space. The encoder unit EU comprises in particular a long short-term memory unit LSTM.

(7) The generator unit GU comprises in particular a decoder unit DU, with said decoder unit DU configured to determine the future trajectory FT of the other road user considering the common embedding space. The decoder unit DU comprises in particular a long short-term memory unit LSTM.

(8) The discriminator unit DU is configured to determine whether the determined future trajectory FT of the other road user is an actual future trajectory of the other road user.

(9) The system comprises in particular an oracle unit OU, with said oracle unit OU configured to determine a reward for the determined future trajectory FT of the at least one other road user considering whether the determined future trajectory FT of the other road user is collision-free.

(10) The system is configured to train the generator unit GU and the discriminator DU unit simultaneously with gradient descent, where said system is in particular configured to train the generator unit GU considering the reward determined by the oracle unit OU.

(11) FIG. 2 shows an exemplified grid in polar angle space. The grid in polar angle space is in particular centered at an other road user for which a future trajectory will be determined by the generator unit GU and aligned with the heading of the other road user. The encoded value per grid cell is in particular the minimum distance to the (static or dynamic) objects o1, o2, o3, o4 that lie within the cone of the angular grid cell.

(12) Compared to a standard 2D grid, the resolution of the grid in polar angle space only influences the dimensionality of the input linearly, while still being able to capture radial distance changes with continuous resolution instead of discretized grid cells. In addition, the observability of angular position changes of surrounding pedestrians becomes more precise the closer they are to the query agent.

(13) The foregoing disclosure has been set forth merely to illustrate the invention and is not intended to be limiting. Since modifications of the disclosed embodiments incorporating the spirit and substance of the invention may occur to persons skilled in the art, the invention should be construed to include everything within the scope of the appended claims and equivalents thereof.