Apparatus and methods for improved road agent behavior prediction using multi-task learning
OwnerAndrei Volchek, Roman Udovichenkoinventors
Life of this patent
Estimating the term…
Abstract
Provided herein is an apparatus and method for improved road agent behavior prediction using multi-task learning, the apparatus including: at least one processor; and a memory, the memory contains instructions configuring the at least one processor to: receive, from one or more vehicles, a real-world road data, wherein the real-world data has been collected using at least in part a data acquisition system of the one or more vehicles; generate, from the real-world data, a ground-truth dataset including ground-truth trajectories, ground-truth intents, and calculated feature values; and train, using the ground-truth dataset, a multi-task machine learning model.
Claims20
An apparatus for improved road agent behavior prediction using multi-task learning, the apparatus comprising:
- at least one processor;
- and
- a memory, wherein the memory is communicatively connected to the at least one processor, the memory contains instructions configuring the at least one processor to:
- receive, from one or more vehicles, a real-world road data, wherein the real-world data has been collected using at least in part a data acquisition system of the one or more vehicles;
- generate, from the real-world data, a ground-truth dataset comprising ground-truth trajectories, ground-truth intents, and calculated feature values;
- and
- train, using the ground-truth dataset, a multi-task machine learning model, wherein training the multi-task machine learning model comprises:
- generating, using the real-world data, a plurality of trajectory predictions using a trajectory model;
- generating, using the real-world data, a plurality of intent predictions using an intent model;
- and
- generating, using the real-world data, a plurality of auxiliary feature predictions using an auxiliary feature model;
- and
- training multi-task machine-learning model using a multi-task loss function, wherein the multi-task loss function is a function of a comparison between the plurality of trajectory predictions and the ground-truth trajectories, a comparison between the plurality of intent predictions and the ground-truth intents, and a comparison between the plurality of auxiliary feature predictions and the calculated feature values.
An method for improved road agent behavior prediction using multi-task learning, the method comprising:
- receiving, using at least one processor and from one or more vehicles, a real-world road data, wherein the real-world data has been collected using at least in part a data acquisition system of the one or more vehicles;
- generating, using the at least one processor and from the real-world data, a ground-truth dataset comprising ground-truth trajectories, ground-truth intents, and calculated feature values;
- and
- training, using the at least one processor and the ground-truth dataset, a multi-task machine learning model, wherein training the multi-task machine learning model comprises:
- generating, using the real-world data, a plurality of trajectory predictions using a trajectory model;
- generating, using the real-world data, a plurality of intent predictions using an intent model;
- and
- generating, using the real-world data, a plurality of auxiliary feature predictions using an auxiliary feature model;
- and
- training multi-task machine-learning model using a multi-task loss function, wherein the multi-task loss function is a function of a comparison between the plurality of trajectory predictions and the ground-truth trajectories, a comparison between the plurality of intent predictions and the ground-truth intents, and a comparison between the plurality of auxiliary feature predictions and the calculated feature values.
Dependencies are inferred from the claim text.