Traffic prediction with reparameterized pushforward policy for autonomous vehicles
View Patent ↗Systems and methods for vehicle behavior prediction include an imaging device that captures images of a vehicle in traffic. A processing device including policy stored in a memory of the processing device in communication with the imaging device stochastically models future behavior of the vehicle based on the captured images. A policy simulator in communication with the processing device simulates the policy as a reparameterized pushforward policy of a base distribution. An evaluator receives the simulated policy from the policy simulator and performs cross-entropy optimization on the future behavior of the vehicle by analyzing the simulated policy and updating the policy according to cross-entropy error. An alert system retrieves the future behavior of the vehicle and recognizes hazardous trajectories of the future trajectories and generates an audible alert using a speaker.
1. A system for vehicle behavior prediction, the system comprising:
an imaging device that captures images of a vehicle in traffic;
a processing device including policy stored in a memory of the processing device in communication with the imaging device to stochastically model future behavior of the vehicle based on the captured images;
a policy simulator in communication with the processing device that simulates the policy as a reparameterized pushforward policy of a base distribution;
an evaluator that receives the simulated policy from the policy simulator and performs cross-entropy optimization on the future behavior of the vehicle by analyzing the simulated policy and updating the policy according to cross-entropy error, the optimization balancing precision and density evaluations for future trajectories by:
min
π
E
ϕ
~
Φ
-
E
x
~
p
(
·
|
ϕ
)
log
q
π
(
x
|
ϕ
)
-
β
E
x
~
q
π
(
·
|
ϕ
)
log
p
(
x
|
ϕ
)
,
where E represents an expected value, ϕ represents a set of features of an environment, Φ represents a ground truth feature distribution, x represents a sample, β represents a predefined coefficient, and p represents a ground truth distribution; and
an alert system that retrieves the future behavior of the vehicle and recognizes hazardous trajectories of the future trajectories and generates an audible alert using a speaker.
2. The system as recited in claim 1 , further including a density estimator that estimates a ground-truth probability for evaluating the future behavior.
3. The system as recited in claim 1 , further including a policy model that simulates the policy as an autoregressive map of random noise sequences in terms of deterministic drift and stochastic diffusion.
4. The system as recited in claim 3 , further including a sampler to sample the random noise sequences from a base distribution.
5. The system as recited in claim 4 , wherein the base distribution is a Gaussian distribution.
6. The system as recited in claim 1 , wherein the evaluator includes the cross-entropy optimization including comparing the future behavior to a probability distribution of an example distribution to determine error of density of a predicted probability distribution corresponding to the future behavior and precision of probabilities of the predicted probability distribution.
7. The system as recited in claim 1 , wherein the policy is a linear generator.
8. The system as recited in claim 1 , wherein the policy is a convolutional neural network.
9. The system as recited in claim 1 , wherein the policy includes:
an encoder convolutional neural network to generate an environmentally reasoned encoding;
an encoder recurrent neural network in parallel with the encoder convolutional neural network to generate a historically reasoned encoding; and
a decoder recurrent neural network to decode a combination of the environmentally reasoned encoding and the historically reasoned encoding.
10. The system as recited in claim 9 , wherein each of the encoder recurrent neural network and the decoder recurrent neural network include gated recurrent units.
11. A system for vehicle behavior prediction, the system comprising:
an imaging device that captures images of a vehicle in traffic;
a processing device including policy stored in a memory of the processing device in communication with the imaging device to stochastically model future behavior of the vehicle based on the captured images;
a policy simulator in communication with the processing device that simulates the policy as a reparameterized pushforward of a base distribution, including a policy model that simulates the policy as an autoregressive map of random noise sequences in terms of deterministic drift and stochastic diffusion;
a density estimator that estimates a ground-truth probability for evaluating the future behavior;
an evaluator in communication with the policy simulator and the density estimator that performs cross-entropy optimization on the future behavior of the vehicle by analyzing the simulated policy and the ground-truth probability according to density of a probability distribution and precision of predicted probabilities corresponding to the future behavior, and updating the policy according to cross-entropy error, the optimization balancing precision and density evaluations for future trajectories by:
min
π
E
ϕ
~
Φ
-
E
x
~
p
(
·
|
ϕ
)
log
q
π
(
x
|
ϕ
)
-
β
E
x
~
q
π
(
·
|
ϕ
)
log
p
(
x
|
ϕ
)
,
where E represents an expected value, ϕ represents a set of features of an environment, Φ represents a ground truth feature distribution, x represents a sample, β represents a predefined coefficient, and p represents a ground truth distribution; and
an alert system that recognizes hazardous trajectories of the probable future trajectories and generates and audible alert using a speaker.
12. The system as recited in claim 11 , further including a sampler to sample the random noise sequences from a base distribution.
13. The system as recited in claim 12 , wherein the base distribution is a Gaussian distribution.
14. The system as recited in claim 11 , wherein the evaluator includes the cross-entropy optimization including comparing the future behavior a probability distribution of an example distribution to determine error of density of a predicted probability distribution corresponding to the future behavior and precision of probabilities of the predicted probability distribution.
15. The system as recited in claim 11 , wherein the policy includes:
an encoder convolutional neural network to generate an environmentally reasoned encoding;
an encoder recurrent neural network in parallel with the encoder convolutional neural network to generate a historically reasoned encoding; and
a decoder recurrent neural network to decode a combination of the environmentally reasoned encoding and the historically reasoned encoding.
16. The system as recited in claim 15 , wherein each of the encoder recurrent neural network and the decoder recurrent neural network include gated recurrent units.
17. A method for vehicle behavior prediction, the method comprising:
capturing images of a vehicle in traffic an imaging device;
stochastically modelling future behavior of the vehicle with policy stored in a memory of a processing device based on the captured images;
simulating the policy as a reparameterized pushforward of a base distribution with a policy simulator;
performing cross-entropy optimization on the future behavior of the vehicle by analyzing the simulated policy and updating the policy according to cross-entropy error using an evaluator, the optimization balancing precision and density evaluations for future trajectories by:
min
π
E
ϕ
~
Φ
-
E
x
~
p
(
·
|
ϕ
)
log
q
π
(
x
|
ϕ
)
-
β
E
x
~
q
π
(
·
|
ϕ
)
log
p
(
x
|
ϕ
)
,
where E represents an expected value, ϕ represents a set of features of an environment, Φ represents a ground truth feature distribution, x represents a sample, β represents a predefined coefficient, and p represents a ground truth distribution; and
recognizing hazardous trajectories of the probable future trajectories and generates and audible alert using a speaker with an alert system.
18. The system as recited in claim 17 , further including estimating a ground-truth probability with a density estimator for evaluating the future behavior.
19. The system as recited in claim 17 , further including simulating the policy as an autoregressive map of random noise sequences in terms of deterministic drift and stochastic diffusion using a policy model.
20. The system as recited in claim 17 , wherein the cross-entropy optimization includes comparing the future behavior to a probability distribution of an example distribution to determine error of density of a predicted probability distribution corresponding to the future behavior and precision of probabilities of the predicted probability distribution.