Overview of the Crowd Analysis Setup

This repository accompanies an academic publication, and an updated, improved version of the Crowdbot dataset. It provides a reproducible analysis pipeline for pedestrian behavior in crowds. By analyzing motion metrics and proxemics, we study differences between human–human interactions (HHI) and human–robot interactions (HRI) in crowded public spaces across CrowdBot, SCAND, JRDB (train and test), and SiT.

---

Files & Folders

The repository layout is as follows (key items):

• AB3DMOT/ — LiDAR-based tracking (package: ab3dmot), original repo: https://github.com/xinshuoweng/AB3DMOT
• checkpoints/ — pre-trained detector weights (e.g., DR-SPAAM, Person_MinkUNet). Uploaded together with the dataset.
• crowd_analysis/
  • crowd_behavior.ipynb — main analysis notebook with all motion metrics and proxemics analysis
• datasets_configs/ — dataset configuration YAMLs
  • data_path_Crowdbot.yaml
  • data_path_JRDB.yaml
  • data_path_SCAND.yaml
  • data_path_SiT.yaml
• datasets_utils/ — dataset utilities (package: crowdbot_data) used by both environments
• lidar_det_2D_3D/ — LiDAR detection (package: lidar_det) combining
  • Person_MinkUNet (3D): https://github.com/VisualComputingInstitute/Person_MinkUNet
  • DR-SPAAM (2D): https://github.com/VisualComputingInstitute/DR-SPAAM-Detector
• rosbags_extraction/ — scripts for ROS bag processing
  • 1_Lidar_from_rosbags.py
  • 2_Pose_from_rosbags.py
  • 3_Detections_from_lidar.py
  • 4_Tracks_from_detections.py
  • Extract_gt_JRDB.py
  • Extract_SiT.py
• run_pipeline.sh — script with a full processing pipeline for generating the input data
• requirements.txt — Python packages installed into crowd_env

---

Dataset

Structure

graph LR;
ROOT[Dataset root];
ROOT-->CK[checkpoints/];
CK-->CKP[*.pth];

ROOT-->RB[rosbags/];
RB-->RBX[_defaced/];
RBX-->BAGS[.bag];

ROOT-->PR[processed/];
PR-->PRX[*_processed/];
PRX-->ALG[alg_res/];
ALG-->DET[detections/];
ALG-->TRK[tracks/];
PRX-->L3D[lidars/];
PRX-->L2D[lidars_2d/];
PRX-->PED[ped_data/];
PRX-->SRC[source_data/];
SRC-->TF[tf_robot/];
SRC-->TS[timestamp/];

Loading

Demo

Item	Preview
(a) Pedestrian trajectories
(b) Motion metrics distributions
(c) Minimum distance distributions
(d) Linear minimum distance vs. robot velocity

---

Proposed/Recommended environment setup with tested package versions

Python version used: 3.8.10

Two Conda environments are used:

• ros_env — ROS I/O from rosbags (bag reading, TF transforms, message types).
• crowd_env — Deep-learning detection/tracking + analysis/visualization.

1) Create ros_env (RoboStack Noetic)

RoboStack brings ROS Noetic into Conda directly (guide: https://robostack.github.io/noetic.html).

mamba create -n ros_env -c conda-forge -c robostack-noetic ros-noetic-desktop ros-noetic-tf2-sensor-msgs
mamba activate ros_env

# Minimal math/transforms used by ros-side scripts
pip install scipy==1.16.2 numpy-quaternion==2024.0.12
pip install python-lzf==0.2.4

SCAND recordings that expose only /velodyne_packets also require velodyne-decoder in ros_env. Recordings containing /velodyne_points do not need it.

RoboStack already provides the compiled message/runtime bits; no extra apt is needed.

2) Create crowd_env

Create the environment (CUDA 11.8 + PyTorch 2.0.0 as tested):

mamba create -n crowd_env   python=3.8.10 ipykernel   cuda-toolkit   pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.8   setuptools=69.5.1   mkl=2023.2.0 mkl-include=2023.2.0 mkl-devel=2023.2.0   -c "nvidia/label/cuda-11.8.0" -c pytorch -c nvidia
mamba activate crowd_env

# Install remaining packages for crowd_env
pip install -r requirements.txt

TorchSparse (install from source — version 2.0.0)

torchsparse==2.0.0 is required for 3D detection and must be installed from source. See the official repository/instructions:
https://github.com/mit-han-lab/torchsparse

Ensure your PyTorch CUDA version is compatible (this setup uses CUDA 11.8 with PyTorch 2.0.0).

Local packages (editable installs)

• Install in both ros_env and crowd_env:

  # Dataset Utils (package: crowdbot_data)
  mamba activate ros_env && pip install -e ./datasets_utils
  mamba activate crowd_env && pip install -e ./datasets_utils

• Install only in crowd_env:

  # LiDAR Detection (package: lidar_det)
  pip install -e ./lidar_det_2D_3D
  # internal libs
  pip install -e ./lidar_det_2D_3D/lib/iou3d
  pip install -e ./lidar_det_2D_3D/lib/jrdb_det3d_eval
  
  # LiDAR-based Tracking (package: ab3dmot) — original repo: https://github.com/xinshuoweng/AB3DMOT
  pip install -e ./AB3DMOT

---

Pipeline overview

The repository provides .ipynb and .py processing scripts. They take as input processed rosbags or prepared LiDAR data from CrowdBot, SCAND, JRDB, and SiT, and produce outputs in a unified CrowdBot data convention for crowd behavior analysis.

Before running anything, edit the YAML files in datasets_configs/. Relative values are resolved from the YAML file's directory (the checked-in examples therefore point to the repository's data/ directory); absolute paths and environment variables are also accepted. Run the commands below from the repository root.

Four processing stages

1. 1_Lidar_from_rosbags.py — Extracts synchronized 2D/3D lidar for CrowdBot, JRDB train, and SCAND. For JRDB test, it reads the released upper/lower PCD streams, timestamps, and SteamLO odometry directly. (uses ros_env)
2. 2_Pose_from_rosbags.py — Extracts and interpolates robot pose for CrowdBot, JRDB, and SCAND; JRDB test poses are loaded from SteamLO CSV files. (uses ros_env)
3. 3_Detections_from_lidar.py — Runs 3D Person-MinkUNet and optional 2D DR-SPAAM detection. JRDB test is 3D-only because its released test set has no matching 2D lidar stream. (uses crowd_env)
4. 4_Tracks_from_detections.py — Builds 3D and optional merged 2D/3D tracks with AB3DMOT. (uses crowd_env)

Dataset-specific extractors

• Extract_gt_JRDB.py — extracts ground truth for JRDB only.
• Extract_SiT.py — extracts LiDAR, egomotion, and labels for SiT.

Full pipeline

Every Python stage has an explicit command-line interface; use --help for all options. The wrapper accepts a dataset, path YAML, logical folder, 3D checkpoint, and optional 2D checkpoint:

bash rosbags_extraction/run_pipeline.sh \
  CrowdBot datasets_configs/data_path_Crowdbot.yaml 0325_rds_defaced \
  checkpoints/ckpt_e40_train_val.pth \
  checkpoints/jrdb_dr_spaam_with_bev_box_e20.pth

For JRDB train, set JRDB_TRAIN_TIMESTAMPS_ROOT. For JRDB test, set both JRDB_TEST_ROOT and JRDB_TEST_ODOM_ROOT; omit the optional 2D model argument. SCAND resolves the Jackal/Spot odometry and lidar topics from each bag automatically.

Revised analysis notebook

Open crowd_analysis/crowd_behavior.ipynb after the tracking outputs exist. Its first parameter cell resolves repository paths, selects one dataset, and documents the final settings. Run the extraction/filtering sections once for each of CROWDBOT, JRDB, SCAND, and SiT; the registration cell retains each dataset's result tables for the cross-dataset plots.

The curated notebook contains the analyses retained in the revised manuscript:

• cubic Savitzky–Golay position smoothing over 1.5 s, followed by finite-horizon differences (1 s velocity, acceleration, and turning; 0.5 s jerk);
• symmetric co-motion exclusion using at least 1 s of shared observations and a 2 m maximum distance excursion, plus the reported threshold sensitivity;
• per-pedestrian Mann–Whitney tests and group-level Wasserstein-1 effect sizes;
• significant-event detection (turn above 45° or speed change above 2 km/h within 1 s) and the joint energy-distance test;
• edge-to-edge close-pass KDE analysis, equalization distance, and maximum excess clearance;
• comfort-zone intrusion frequencies, density stratification on 4 s fragments, and robot/pedestrian speed controls using robust LOWESS (frac=0.3, two robust iterations).

---

References:

Please cite both the dataset as well as the publication if you use our dataset/repository in your work.

Crowdbot_v2 dataset

Wojcikiewicz, D., Billard, A., & Paez-Granados, D. (2025). CrowdBot_v2: Pedestrian–Robot crowd navigation dataset with pedestrian tracking (v2.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17694140

Academic Publication

Wojcikiewicz D., Billard A., Paez-Granados D. Assessing pedestrian responses to autonomous and personal mobility robots in crowded public spaces. Science Advances (2026). https://doi.org/10.1126/sciadv.aef2576

---

Acknowledgment

This research work was partially supported by the Innosuisse Project 103.421 IP-IC "Developing an AI-enabled Robotic Personal Vehicle for Reduced Mobility Population in Complex Environments" and the JST Moonshot R&D [Grant Number JPMJMS2034-18].