Real-time Crowd Segmentation from 2D/3D lidar person detections

Simulating crowds of people in Isaac Sim

This repository...

If you find this code relevant for your work, please consider citing one or both of these papers. A bibtex entry is provided below:

@article{hughes2023foundations,
         title={TODO},
         author={TODO},
         year={2024},
         eprint={xxxx.xxxxx},
         archivePrefix={arXiv},
         primaryClass={cs.RO}
}

Installation

Requirements

• GPU for machine learning
  • When not using CUDA install pytorch manually!

Setup

1. Create a virtual python environment with python3 -m venv env
2. Enter into the environment using source env/bin/activate
3. Install required packages pip install -r requirements.txt

Configuration

Most basic configurations for both input/label generation and the scene segmentation itself can be found in the config folder.

Run

Input and Label Generator

Requirements

Have detection data in either CrowdBot or JRDB (with additional velocities of all people) format. See this repository for a simulator to create data in JRDB format.

Usage with JRDB based data

Run python3 generator/generator.py --dataformat "JRDB" with the following additional arguments.

• --input with the path to your detection data
• --output with the path where the generated images should be stored.

Usage with CrowdBot data

This requires a few more arguments

Run python3 generator/generator.py --dataformat "CrowdBot" with the following additional arguments.

• --input with the path to your detection data
• --output with the path where the generated images should be stored.
• --input_tf with the path to the file with all tf frames.
• --input_vel with the path to the file with all detection velocities.

Additionally the argument --label "False" can be added for both cases to only generate the input without generating all the label.

Scene segmentation

Train

1. Create a text file at the location given by data_root in the config/config.yml file. This file should have all relative paths to the training data folder.

Optionally: The same can be done for the validation data using validation_dir as the config variable

2. Run python3 engine/train.py to start training of the model.

Predict

Run python3 engine/predict.py with the following arguments.

• --input with the location to the input images you want to predict
• --output with the location where the predicted segmentations should be stored
• --checkpoint with the path to the checkpoint file to restart at. If empty it will retrain with the given data.