Systems and methods for predicting the trajectory of a road agent external to a vehicle
Systems and methods described herein relate to predicting a trajectory of a road agent external to a vehicle. One embodiment generates first and second predicted road-agent trajectories using respective first and second trajectory predictors based, at least in part, on a plurality of inputs including past road-agent trajectory information and vehicle sensor data; generates a confidence score for each predicted road-agent trajectory using a confidence estimator that includes a deep neural network, wherein generating the confidence scores includes computing them as a function of time within a predetermined temporal horizon; outputs the first and second predicted road-agent trajectories and their respective confidence scores; and controls operation of the vehicle based, at least in part, on one or more of the first predicted road-agent trajectory, the second predicted road-agent trajectory, the confidence score for the first predicted road-agent trajectory, and the confidence score for the second predicted road-agent trajectory.
1. A system for predicting a trajectory of a road agent, the system comprising:
one or more processors; and
a memory communicably coupled to the one or more processors and storing:
a trajectory-prediction module including instructions that when executed by the one or more processors cause the one or more processors to:
generate a first predicted road-agent trajectory using a first trajectory predictor, wherein the road agent is external to a vehicle and the first trajectory predictor generates the first predicted road-agent trajectory based, at least in part, on a plurality of inputs that include past road-agent trajectory information and information from one or more sensors of the vehicle;
generate a second predicted road-agent trajectory using a second trajectory predictor, wherein the second trajectory predictor generates the second predicted road-agent trajectory based, at least in part, on the plurality of inputs; and
generate a confidence score for each of the first and second predicted road-agent trajectories using a confidence estimator that includes a first deep neural network, wherein the confidence scores are computed as a function of time within a predetermined temporal horizon;
a trajectory-output module including instructions that when executed by the one or more processors cause the one or more processors to output the first and second predicted road-agent trajectories and their respective confidence scores; and
a control module including instructions that when executed by the one or more processors cause the one or more processors to control operation of the vehicle based, at least in part, on one or more of the first predicted road-agent trajectory, the second predicted road-agent trajectory, the confidence score for the first predicted road-agent trajectory, and the confidence score for the second predicted road-agent trajectory.
2. The system of claim 1 , wherein the road agent is one of an automobile, a motorcycle, a bicycle, and a pedestrian.
3. The system of claim 1 , wherein the first trajectory predictor is a probabilistic variational trajectory predictor that includes a second deep neural network.
4. The system of claim 3 , wherein the trajectory-prediction module includes instructions to generate the first predicted road-agent trajectory, at least in part, by computing Gaussian mixture model (GMM) parameters for one or more projection coefficients of a basis function.
5. The system of claim 4 , wherein the GMM parameters include at least one of a weight, a mean, and a variance.
6. The system of claim 3 , further comprising:
a model-training module including instructions that when executed by the one or more processors cause the one or more processors to train the first and second deep neural networks.
7. The system of claim 1 , wherein the past road-agent trajectory information includes a sequence of spatial coordinates projected onto one or more coefficients of a basis function.
8. The system of claim 1 , wherein the information from one or more sensors of the vehicle includes at least one of image data, light detection and ranging (LIDAR) data, radar data, and sonar data.
9. The system of claim 1 , wherein the second trajectory predictor is a physics-based model.
10. The system of claim 1 , wherein the first and second trajectory predictors use different computational models to generate the respective first and second predicted road-agent trajectories.
11. The system of claim 1 , wherein the trajectory-output module includes further instructions to output, as a most likely predicted trajectory of the road agent, the one of the first and second predicted road-agent trajectories whose confidence score corresponds to a higher level of confidence, and the control module includes instructions to plan a trajectory for the vehicle based, at least in part, on the most likely predicted trajectory of the road agent.
12. A non-transitory computer-readable medium for predicting a trajectory of a road agent and storing instructions that when executed by one or more processors cause the one or more processors to:
generate a first predicted road-agent trajectory using a first trajectory predictor, wherein the road agent is external to a vehicle and the first trajectory predictor generates the first predicted road-agent trajectory based, at least in part, on a plurality of inputs that include past road-agent trajectory information and information from one or more sensors of the vehicle;
generate a second predicted road-agent trajectory using a second trajectory predictor, wherein the second trajectory predictor generates the second predicted road-agent trajectory based, at least in part, on the plurality of inputs;
generate a confidence score for each of the first and second predicted road-agent trajectories using a confidence estimator that includes a first deep neural network, wherein the confidence scores are computed as a function of time within a predetermined temporal horizon;
output the first and second predicted road-agent trajectories and their respective confidence scores; and
control operation of the vehicle based, at least in part, on one or more of the first predicted road-agent trajectory, the second predicted road-agent trajectory, the confidence score for the first predicted road-agent trajectory, and the confidence score for the second predicted road-agent trajectory.
13. The non-transitory computer-readable medium of claim 12 , wherein:
the first trajectory predictor is a probabilistic variational trajectory predictor that includes a second deep neural network;
the second trajectory predictor is a physics-based model; and
the instructions to generate the first predicted road-agent trajectory include instructions to compute Gaussian mixture model (GMM) parameters for one or more projection coefficients of a basis function, the GMM parameters including at least one of a weight, a mean, and a variance.
14. A method of predicting a trajectory of a road agent, the method comprising:
generating a first predicted road-agent trajectory using a first trajectory predictor, wherein the road agent is external to a vehicle and the first trajectory predictor generates the first predicted road-agent trajectory based, at least in part, on a plurality of inputs that include past road-agent trajectory information and information from one or more sensors of the vehicle;
generating a second predicted road-agent trajectory using a second trajectory predictor, wherein the second trajectory predictor generates the second predicted road-agent trajectory based, at least in part, on the plurality of inputs;
generating a confidence score for each of the first and second predicted road-agent trajectories using a confidence estimator that includes a first deep neural network, wherein generating the confidence scores includes computing the confidence scores as a function of time within a predetermined temporal horizon;
outputting the first and second predicted road-agent trajectories and their respective confidence scores; and
controlling operation of the vehicle based, at least in part, on one or more of the first predicted road-agent trajectory, the second predicted road-agent trajectory, the confidence score for the first predicted road-agent trajectory, and the confidence score for the second predicted road-agent trajectory.
15. The method of claim 14 , wherein the road agent is one of an automobile, a motorcycle, a bicycle, and a pedestrian.
16. The method of claim 14 , wherein the first trajectory predictor generates the first predicted road-agent trajectory in accordance with a probabilistic variational model that includes a second deep neural network.
17. The method of claim 16 , wherein generating the first predicted road-agent trajectory includes computing Gaussian mixture model (GMM) parameters for one or more projection coefficients of a basis function, the GMM parameters including at least one of a weight, a mean, and a variance.
18. The method of claim 16 , further comprising training the first and second deep neural networks.
19. The method of claim 14 , wherein the first and second trajectory predictors generate the respective first and second predicted road-agent trajectories in accordance with different computational models.
20. The method of claim 14 , further comprising:
outputting a most likely predicted trajectory of the road agent by selecting the one of the first and second predicted road-agent trajectories whose confidence score corresponds to a higher level of confidence; and
planning a trajectory for the vehicle based, at least in part, on the most likely predicted trajectory of the road agent.