Methods and systems for diversity-aware vehicle motion prediction via latent semantic sampling
A system and method for generating a predicted vehicle trajectory includes a generative adversarial network configured to receive a trajectory vector of a target vehicle and generate a set of latent state vectors using the received trajectory vector and an artificial neural network. The latent state vectors each comprise a high-level sub-vector, Z H . The GAN enforces Z H to be correlated to an annotation coding representing semantic categories of vehicle trajectories. The GAN selects a subset, from the set of latent state vectors, using farthest point sampling and generates a predicted vehicle trajectory based on the selected subset of latent state vectors.
1. A method for generating a predicted vehicle trajectory comprising:
receiving a trajectory vector of a target vehicle;
generating a set of latent state vectors using the received trajectory vector and an artificial neural network, wherein the latent state vectors each comprise a high-level representation, Z H , correlated to an annotation coding representing semantic categories of vehicle trajectories;
selecting a subset, from the set of latent state vectors, using farthest point sampling;
generating a predicted vehicle trajectory based on the subset of latent state vectors; and
controlling a vehicle based on the predicted vehicle trajectory.
2. The method of claim 1 , further comprising:
selecting the subset, from the set of latent state vectors, using farthest point sampling on the high-level representation, Z H ;
weighting each latent state vector in the subset with a Voronoi cell weight; and
generating the predicted vehicle trajectory based on the weighted subset.
3. The method of claim 2 , wherein the generating of the set of latent state vectors comprises:
receiving a map vector comprising at least one point corresponding to a road lane;
generating an intermediate vector using a first recurrent neural network on the trajectory vector and the map vector;
generating a noise vector from a normal distribution; and
generating a latent state vector using a linear layer neural network on the intermediate vector and the noise vector.
4. The method of claim 3 , further comprising:
generating the predicted vehicle trajectory using a second recurrent neural network on the map vector and a weighted latent state vector in the subset.
5. The method of claim 3 , wherein the latent state vectors each further comprise a low-level representation, Z L ; and
wherein the generating of the latent state vector comprises encouraging the high-level representation, Z H , and the low-level representation, Z L , to be uncorrelated.
6. The method of claim 5 , further comprising:
encouraging the high-level representation, Z H , to be correlated with the annotation coding such that differences between latent state vectors associated with the same annotation coding are smaller than differences between latent state vectors associated with different annotation codings.
7. The method of claim 6 , wherein the encouraging of the high-level representation, Z H , and the low-level representation, Z L , to be uncorrelated comprises using a latent space regularization loss;
wherein the encouraging of the high-level representation, Z H , to be correlated with the annotation coding comprises using an embedding loss; and
using a discriminator to generate a loss signal for the generated predicted vehicle trajectory, wherein the loss signal comprises the latent space regularization loss and the embedding loss.
8. A non-transitory computer-readable storage medium comprising instructions that, when executed by a processor, cause the processor to perform the steps of:
receiving a trajectory vector of a target vehicle;
generating a set of latent state vectors using the received trajectory vector and an artificial neural network, wherein the latent state vectors each comprise a high-level representation, Z H , correlated to an annotation coding representing semantic categories of vehicle trajectories;
selecting a subset, from the set of latent state vectors, using farthest point sampling;
generating a predicted vehicle trajectory based on the subset of latent state vectors; and
controlling a vehicle based on the predicted vehicle trajectory.
9. The non-transitory computer-readable storage medium of claim 8 , further comprising instructions that when executed by a processor, cause the processor to further perform steps comprising:
selecting the subset, from the set of latent state vectors, using farthest point sampling on the high-level representation, ZH;
weighting each latent state vector in the subset with a Voronoi cell weight; and
generating the predicted vehicle trajectory based on the weighted subset.
10. The non-transitory computer-readable storage medium of claim 9 , wherein the generating of the set of latent state vectors comprises:
receiving a map vector comprising at least one point corresponding to a road lane;
generating an intermediate vector using a first recurrent neural network on the trajectory vector and the map vector;
generating a noise vector from a normal distribution; and
generating a latent state vector using a linear layer neural network on the intermediate vector and the noise vector.
11. The non-transitory computer-readable storage medium of claim 10 , further comprising instructions that when executed by a processor, cause the processor to further perform steps comprising:
generating the predicted vehicle trajectory using a second recurrent neural network on the map vector and a weighted latent state vector in the subset.
12. The non-transitory computer-readable storage medium of claim 10 , wherein the latent state vectors each further comprise a low-level representation, Z L ; and
wherein the generating of the latent state vector comprises encouraging the high-level representation, Z H , and the low-level representation, Z L , to be uncorrelated.
13. The non-transitory computer-readable storage medium of claim 12 , further comprising instructions that when executed by a processor, cause the processor to further perform steps comprising:
encouraging the high-level representation, Z H , to be correlated with the annotation coding such that differences between latent state vectors associated with the same annotation coding are smaller than differences between latent state vectors associated with different annotation codings.
14. The non-transitory computer-readable storage medium of claim 13 , further comprising instructions that when executed by a processor, cause the processor to further perform steps comprising:
using a latent space regularization loss to encourage the high-level representation, Z H , and the low-level representation, Z L , to be uncorrelated;
using an embedding loss to encourage the high-level representation, Z H , to be correlated with the annotation coding; and
using a discriminator to generate a loss signal for the generated predicted vehicle trajectory, wherein the loss signal comprises the latent space regularization loss and the embedding loss.
15. A computing device comprising:
a processor; and
a memory storing instructions that, when executed by the processor, cause the computing device to:
receive a trajectory vector of a target vehicle;
generate a set of latent state vectors using the received trajectory vector and an artificial neural network, wherein the latent state vectors each comprise a high-level representation, Z H , correlated to an annotation coding representing semantic categories of vehicle trajectories;
select a subset, from the set of latent state vectors, using farthest point sampling;
generate a predicted vehicle trajectory based on the subset of latent state vectors; and
control a vehicle based on the predicted vehicle trajectory.
16. The computing device of claim 15 , wherein the memory further comprises instructions that, when executed by the processor, further cause the computing device to:
select the subset, from the set of latent state vectors, using farthest point sampling on the high-level representation, Z H ;
weight each latent state vector in the subset with a Voronoi cell weight; and
generate the predicted vehicle trajectory based on the weighted subset.
17. The computing device of claim 16 , wherein the generating of the set of latent state vectors comprises:
receiving a map vector comprising at least one point corresponding to a road lane;
generating an intermediate vector using a first recurrent neural network on the trajectory vector and the map vector;
generating a noise vector from a normal distribution; and
generating a latent state vector using a linear layer neural network on the intermediate vector and the noise vector.
18. The computing device of claim 17 , wherein the non-transitory computer-readable storage medium further comprises instructions that, when executed by the processor, further cause the computing device to:
generate the predicted vehicle trajectory using a second recurrent neural network on the map vector and a weighted latent state vector in the subset.
19. The computing device of claim 17 , wherein the latent state vectors each further comprise a low-level representation, Z L ; and
wherein the generating of the latent state vector comprises encouraging the high-level representation, Z H , and the low-level representation, Z L , to be uncorrelated.
20. The computing device of claim 19 , wherein the non-transitory computer-readable storage medium further comprises instructions that, when executed by the processor, further cause the computing device to:
encourage the high-level representation, Z H , to be correlated with the annotation coding such that differences between latent state vectors associated with the same annotation coding are smaller than differences between latent state vectors associated with different annotation codings.