Future trajectory prediction in multi-actor environment for autonomous machine applications

Abstract

In various examples, past location information and map information corresponding to an active person in an environment may be applied to a deep neural network (DNN), such as a recurrent neural network (RNN), trained to calculate information corresponding to a future trajectory of the active person. The output of the DNN may include, for each future time slice trained to predict the DNN, a confidence map representing a confidence of each pixel in which the active person is present, and a vector field representing a location of the active person in the confidence map of the previous time slice. The vector field may thus be used to track an object through a confidence map for each future time slice to generate a predicted future trajectory for each actor. In addition to the tracked past trajectory, the predicted future trajectory may also be used to generate a complete trajectory for an active person who may assist in the own vehicle while navigating the environment.

Description

Background technique [0001] In order for autonomous vehicles to navigate effectively, autonomous vehicles need to generate an understanding of their surroundings. For example, identifying the location of nearby cars, pedestrians, traffic signs and signals, and road configurations are key aspects for safe control of autonomous vehicles. In addition to the current position and configuration of objects in the environment, determining the likely future trajectories of these objects over time by observation by an autonomous vehicle can prove effective in fully understanding and interpreting predicted changes in the environment. [0002] Conventional systems exploit past object trajectory information using a single actor approach, where an understanding of actors around an ego-vehicle can be determined, and then determinations of future trajectories for each actor can be computed individually. For example, some conventional systems rely on combinatorial networks to determine high-l...

AI Technical Summary

Problems solved by technology

Consequently, these conventional approaches are limited to predicting the future position of a single object at a time (thereby limiting the usefulness of the information for autonomous driving applications) or to predicting many possible trajectories that do not directly correspond to the actual predicted future trajectories of the agent.

With these regular processes repeated for each object, the runtime of the system becomes ineffective for real-time deployment due to the processing burden on the system

Images