An end-to-end agent-map interaction framework for multi-agent trajectory prediction

Li, Xiang

Permalink

https://hdl.handle.net/2142/120114 Copy

Description

Title

An end-to-end agent-map interaction framework for multi-agent trajectory prediction

Author(s)

Li, Xiang

Issue Date

2023-04-27

Director of Research (if dissertation) or Advisor (if thesis)

Driggs-Campbell, Katie

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Autonomous Driving
• Machine Learning
• Trajectory Prediction
• Artificial Intelligence.

Language

eng

Abstract

This thesis proposes an end-to-end agent-map interaction framework for multi-agent trajectory prediction, which can be used in motion prediction tasks in relatively complex traffic scenarios that involve multiple agents, such as vehicles and pedestrians. The framework consists of an end-to-end deep learning model which takes motion history information of agents and rasterized context map image as input and predicts the future trajectories of all agents and occupancy on the map. The agents and map input features are extracted by capturing the interactions among agents, the temporal relationship of agents' motion, and the spatial information contained in the map. Agents and map embeddings interact with each other via a symmetric transformer structure based on cross-attention, then predictions of trajectories and occupancy are generated from post-interaction embeddings. The proposed framework does not utilize any extra internal levels of input representations such as lane graphs, motion heat maps, and waypoints; thus it is easier to be generalized for inputs with different modalities. The framework is implemented and then trained and evaluated on the nuScenes dataset. The framework is compared with many state-of-the-art works

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120114

Copyright and License Information

Copyright 2023 Xiang Li

Owning Collections