ASU Pedestrian LiDAR Scenes (APLS) Dataset

Introduction

Manifest files for APLS-1, APLS-2, APLS-3, and stats_file.csv

Welcome to the ASU Pedestrian LiDAR Scenes (APLS) Dataset. Currently, the dataset is being compressed for upload. A sample can be downloaded here (APLS-sample)

We have collected the APLS dataset to facilitate efforts in point cloud representation learning using a semi-supervised "next frame given previous frames" prediction task.

Description

APLS contains 30 hours of Velodyne HDL-32e point cloud data. It was captured in 20 collection runs in 10 different pedestrian walking areas around the campus of Arizona State University in 2019. The dataset contains ~1 million LiDAR frames.

Table 1: Collection run details
Repo #Run NameLocationTime of DayWeatherTraffic LevelN FramesDuration
1Bookstore-1 Bookstore Afternoon Sunny High 72939 2 hr, 1 min
1Bookstore-2 Bookstore Afternoon Sunny High 39130 1 hr, 5 min
1Church-1 Church Morning Sunny Moderate 58009 1 hr, 36 min
1Church-2 Church Afternoon Sunny Moderate 58793 1 hr, 38 min
1Courtyard Courtyard Afternoon Sunny Sparse 92785 2 hr, 34 min
1Thoroughfare Thoroughfare Mid-day Sunny Moderate 53919 1 hr, 29 min
1Pavilion-1 Pavilion Morning Sunny Moderate 55541 1 hr, 33 min
2Pavilion-2 Pavilion Afternoon Sunny Moderate 54148 1 hr, 30 min
2Gymnasium-1 Gymnasium Sunset & Evening Night High 55613 1 hr, 32 min
2Gymnasium-2 Gymnasium Sunset & Evening Night High 6727 0 hr, 11 min
2Gymnasium-3 Gymnasium Sunset & Evening Night High 45305 1 hr, 15 min
2Intersection-1 Intersection Morning Sunny Moderate 56159 1 hr, 33 min
2Intersection-2 Intersection Afternoon Overcast Moderate 53684 1 hr, 29 min
3Boulevard Boulevard Morning Sunny Low 55284 1 hr, 32 min
3Fountain-1 Fountain Morning Sunny High 52111 1 hr, 27 min
3Fountain-2 Fountain Afternoon Overcast Moderate 24714 0 hr, 41 min
3Bench-1 Bench Morning Sunny Moderate 54117 1 hr, 30 min
3Bench-2 Bench Afternoon Sunny Moderate 55512 1 hr 32 min
3Bench-3 Bench Evening Night Moderate 52110 1 hr, 27 min
3Bench-4 Bench Afternoon & Evening Sunset/Night Moderate 107714 3 hr, 0 min

Figure 1: Campus collection locations


Figure 2: 3D sample of a frame

Figure 3: "Depth camera" sample of a frame

Technical Details

For space saving purposes, the data is stored in "Distance-only Velodyne" (DOVe) format. The files contain only the parts of the captured LiDAR bytestream that contain actual distance data, and all other data is discarded. Each frame is stored as a single file, rather than stored as a chunk of bytes in a package capture file containing all frames from a given run, as the Velodyne PCAP output data is initially stored. Thus, in addition to saving space, it enables in-place arbitrary reading of any frame in the dataset in O(1) without having to advance past all prior frames in time for a run.

The Velodyne Dataset Utils repository contains the file "dove_dataset.py," which defines a Pytorch dataloader (DOVeDataset class) for a directory of DOVe format binary frames. The repo contains usage examples on loading a DOVe dataset such as APLS. Finally, the repo contains "parser.py," which can be used to convert any Velodyne HDLE-32e PCAP file into a DOVe dataset. We will extend the functionality of parser.py to support other LiDAR devices if there is interest.

Usage

A sample may be downloaded from APLS-sample and unzipped using the following:

$cd APLS-sample/
$unzip Gym-2.zip
$cd ..

The dataset can be initialized as Pytorch datasets using the following code, for example for the directory containing "APLS-sample":

1import dove_dataset
2from torch.utils.data import DataLoader
3
4d_train, d_val, d_test = dove_dataset.CreateDOVeDatasets('./APLS-sample/', frames_per_clip=10, raw=False, granularity=3)
5loader_train = DataLoader(d_train, batch_size=32, shuffle=True, num_workers=3)
6for minibatch in d_train:
7     ...