Dynamic scene prediction with multiple interacting agents
Methods and systems for predicting a trajectory include determining prediction samples for agents in a scene based on a past trajectory. The prediction samples are ranked according to a likelihood score that incorporates interactions between agents and semantic scene context. The prediction samples are iteratively refined using a regression function that accumulates scene context and agent interactions across iterations. A response activity is triggered when the prediction samples satisfy a predetermined condition.
1. A method for predicting a trajectory, comprising:
determining a plurality of prediction samples for each agent in a plurality of agents in a scene based on a past trajectory;
ranking the prediction samples using a processor according to a likelihood score that incorporates interactions between agents and semantic scene context;
iteratively refining the prediction samples using a regression function that accumulates scene context and agent interactions across iterations, wherein the regression function is defined as:
η( Ŷ i (k) ; ,X,Ŷ j\i ∀ )
where Ŷ i (k) is an i th prediction sample, is a scene context, X is a set of known past trajectories for the plurality of agents, and Ŷ j\i ∀ is the prediction samples of agents other than an agent I; and
triggering a response activity when the prediction samples satisfy a predetermined condition.
2. The method of claim 1 , further comprising training a neural network used to generate the prediction samples based on past trajectories and outcomes of the past trajectories.
3. The method of claim 2 , wherein training the neural network comprises determining stochastic latent variables that characterize a distribution of prediction samples.
4. The method of claim 3 , wherein determining the plurality of prediction samples comprises randomly generating predictions within parameters set by the stochastic latent variables.
5. The method of claim 3 , wherein a distribution of the stochastic latent variables comprises a Gaussian distribution defined by a mean and a standard deviation, regularized by a Kullback-Leibler divergence over a distribution of the past trajectories and the outcomes of the past trajectories.
6. The method of claim 1 , wherein the agent interactions are characterized by all past trajectories, including trajectories from training data and trajectories in a present scene.
7. The method of claim 1 , wherein the scene context is characterized by static features of a scene that influence agent behavior.
8. The method of claim 1 , wherein the scene is a video feed and wherein agents are moving entities within the scene.
9. The method of claim 1 , wherein the response activity comprises an activity selected from the group consisting of raising an alarm and enabling a barrier to a user's trajectory.
10. A system for predicting a trajectory, comprising:
a prediction sample module configured to determine a plurality of prediction samples for each agent in a plurality of agents in a scene based on a past trajectory;
a ranking/refinement module comprising a processor configured to rank the prediction samples according to a likelihood score that incorporates interactions between agents and semantic scene context and to iteratively refine the prediction samples using a regression function that accumulates scene context and agent interactions across iterations, wherein the regression function is defined as:
η( Ŷ i (k) ; ,X,Ŷ j\i ∀ )
where Ŷ i (k) is an i th prediction sample, is a scene context, X is a set of known past trajectories for the plurality of agents, and Ŷ j\i ∀ is the prediction samples of agents other than an agent I; and
a response module configured to trigger a response activity when the prediction samples satisfy a predetermined condition.
11. The system of claim 10 , further comprising a training module configured to train a neural network used to generate the prediction samples based on past trajectories and outcomes of the past trajectories.
12. The system of claim 11 , wherein the training module is further configured to determine stochastic latent variables that characterize a distribution of prediction samples.
13. The system of claim 12 , wherein the prediction sample module is further configured to randomly generate predictions within parameters set by the stochastic latent variables.
14. The system of claim 12 , wherein a distribution of the stochastic latent variables comprises a Gaussian distribution defined by a mean and a standard deviation, regularized by a Kullback-Leibler divergence over a distribution of the past trajectories and the outcomes of the past trajectories.
15. The system of claim 11 , wherein the agent interactions are characterized by all past trajectories, including trajectories from training data and trajectories in a present scene.
16. The system of claim 11 , wherein the scene context is characterized by static features of a scene that influence agent behavior.
17. The system of claim 11 , wherein the scene is a video feed and wherein agents are moving entities within the scene.
18. The system of claim 11 , wherein the response activity comprises an activity selected from the group consisting of raising an alarm and enabling a barrier to a user's trajectory.