Dragon Lake Parking (DLP) Dataset
What is this?
The Dragon Lake Parking (DLP) Dataset contains annotated video and data of vehicles, cyclists, and pedestrians inside a parking lot. We collected it by flying a drone above a huge parking lot.
Abundant vehicle parking maneuvers and interactions are recorded. To the best of our knowledge, this is the first and largest public dataset designated for the parking scenario (up to April 2022), featuring high data accuracy and a rich variety of realistic human driving behavior.
Annotated Video
Semantic Visualization
Statistics
Raw video
Length: 3.5 hours
Resolution: 4K
Frame rate: 25 fps
Parking Area
Size: 140 m x 80 m
Number of spots: ~400
Agent Types and Count
Vehicles (normal sedan, medium vehicle like SUV, bus): 1216
Pedestrians: 3904
Bicycles: 28
Motorcycles: 5
Data Structure
The raw videos are annotated and converted to JSON format. The dataset has a graph structure with the following components
Agent: An agent is an object that has moved in this scene. It contains the object's dimension, type, and trajectory as a list of instances.
Instance: An instance is the state of an agent at a time step, which includes position, orientation, velocity, and acceleration. It also points to the preceding / subsequent instance along the agent's trajectory.
Frame: A frame is a discrete sample from the recording. It contains a list of visible instances at this time step, and points to the preceding / subsequent frame.
Obstacle: Obstacles are vehicles that never move in this recording.
Scene: A scene represents a consecutive video recording with certain length. It points to all frames, agents, and obstacles in this recording.
The entire DLP dataset contains 30 scenes, 317,873 frames, 5,188 agents, and 15,383,737 instances.
Formats
Two types of data are available:
JSONs will provide you all Instances, Agents, Frames, Obstacles, and Scene data so that you can use our Python toolkit. All coordinates are transformed from UTM to the local coordinates of parking lot. We STRONGLY RECOMMEND that you just use JSONs if your research is about trajectory analysis and/or semantic visualization is enough for your computer vision module.
Raw video and ground truth annotation. You ONLY need this if you are working on object detection, tracking, semantic segmentation, or end2end model with raw bird's eye view camera data. The annotated trajectories are in UTM coordinates. Please understand that we cannot offer software tools for parsing these data.
Citation
The dataset is formally released in this paper by supporting the research of vehicle intent and motion prediction. Please cite the following if the DLP dataset or its Python toolkit is used.
@INPROCEEDINGS{9922162,
author={Shen, Xu and Lacayo, Matthew and Guggilla, Nidhir and Borrelli, Francesco},
booktitle={2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC)},
title={ParkPredict+: Multimodal Intent and Motion Prediction for Vehicles in Parking Lots with CNN and Transformer},
year={2022},
volume={},
number={},
pages={3999-4004},
doi={10.1109/ITSC55140.2022.9922162}}
Python Toolkit
We are releasing a Python toolkit, which provides convenient APIs to query and visualize data.
Download
JSON Data
You can download JSON data files directly with Dryad:
Raw video and ground truth annotation
Since the video & annotation has 168GB in size, please try the sample data before requesting the full dataset.
After trying the sample data, if you would like to request access to the entire dataset, please fill out the form below. Make sure you have clearly specified your reason forrequesting raw video and annotation based on your trial with sample data. Otherwise, your application might be rejected.
Term and Conditions of Usage
By requesting this dataset, you agree to the following term and conditions of usage:
Any use of this dataset should properly cite our publications on the website. https://sites.google.com/berkeley.edu/dlp-dataset
The dataset is used only for non-commercial purposes, including research, teaching, scientific publication and personal experimentation. Non-commercial Purposes include the use of the Dataset to perform benchmarking for purposes of academic or applied research publication. Non-commercial Purposes do not include purposes primarily intended for or directed towards commercial advantage or monetary compensation, or purposes intended for or directed towards litigation, licensing, or enforcement, even in part.
You do not distribute the dataset or any of its modified versions to other individuals, institutes, companies, associations or the public. Any 3rd party should be directed to us and send another request.
Authors
Xu Shen, Michelle Pan, Vijay Govindarajan, Neelay Velingker, Alex Wong, Yibin Li
Please contact us if you have any questions.
Model Predictive Control (MPC) Lab at UC Berkeley
Source Trajectory and Bounding box data were annotated and gathered with DataFromSky TrafficSurvey - an AI video analytics-based service for gathering advanced traffic data.