Surrounding Motion Predict Model for Autonomous Vehicle

Wenhao Cui, Guangrui Shen, Tieming Sun

EE 599 Deep Learning - Fall 2020

Introduction

• Predict surrounding agents motions of the autonomous vehicle over 5s given their historical 1s positions
• Useful for planning self driving vehicle’s movement
• Deep learning techniques (CNN: Mixnet) + Ensemble Models
• Choose negative multi-log-likelihood as evaluate metric
• Full Information provided by Kaggle

Run Model

• Follow the instruction on Lyft Website to download Dataset
• Use Jupyter Notebooks under directory "notebook" to run our model
• Or run python script under "code", first changing your path to dataset
• Structure of this repo

- code - train.py
      |
       - test.py
      |
       - model.py
      |
       - utils.py

- data_model - pth
            |
             - metric

- notebook - train-cnn-nll.ipynb
          |
           - test-cnn.ipynb

Detailed Report

• Final report and presentation are provided under directory "report"

Lyft Prediction Dataset

@misc{lyft2020,
title = {One Thousand and One Hours: Self-driving Motion Prediction Dataset},
author = {Houston, J. and Zuidhof, G. and Bergamini, L. and Ye, Y. and Jain, A. and Omari, S. and Iglovikov, V. and Ondruska, P.},
year = {2020},
howpublished = {\url{https://level5.lyft.com/dataset/}} }