Bayesian Floor Field

Authors: Francesco Verdoja (francesco.verdoja@aalto.fi), Tomasz Piotr Kucner, and Ville Kyrki
Affiliation: School of Electrical Engineering, Aalto University, Finland

Bayesian Floor Field (BFF) is a novel approach to learn people flow able to combine knowledge about the environment geometry with observations from human trajectories. An occupancy-based deep prior is used to build an initial transition model without requiring any observations of pedestrians; the model is then updated when observations become available using Bayesian inference. We demonstrate the ability of our model to increase data efficiency and to generalize across real large-scale environments, which is unprecedented for maps of dynamics.

Presentation

Installation

Software requirements

• Python 3.9
• PyTorch 1.9.1
• CUDA 11.8

A complete list of dependencies is listed in requirements.txt and should be installed for the software to run.
If you are interested in running the tests or doing development, you should also install the dependencies in requirements-dev.txt.

Submodules

Our Tiramisu network implementation is based on OctoPyTorch. We incorporate OctoPyTorch as git submodule. To pull the submodule, after cloning this project, you just have to run:

git submodule update --init --recursive

Reproducing the results of the paper

Data

Link to download the data and models (695 MB)

Data structure

BFF_data_models.zip
│   LICENSE  # MIT licence
│   README.md  # data description
│
├───data  # folder containing datasets
│   ├───ATC  # ATC dataset
│   │   │   localization_grid.pgm  # occupancy map image
│   │   │   localization_grid.yaml  # occupancy map metadata
│   │   │
│   │   └───subsampled
│   │           atc-20121114_5.csv  # curated data from 11/14/2012 (training)
│   │           atc-20121118_5.csv  # curated data from 11/18/2012 (testing)
│   │
│   └───KTH_track  # KTH dataset
│           kth_trajectory_data.csv  # curated data, consistent with ATC format
│           map.pgm  # occupancy map image
│           map.yaml  # occupancy map metadata
│
├───maps  # occupancy-based prior maps
│       map_atc_w64_s8_t_120.npy
│       map_atc_w64_s16_t_120.npy
│       map_atc_w64_s20_t_120.npy
│       map_kth_w64_s8_t_120.npy
│       map_kth_w64_s16_t_120.npy
│       map_kth_w64_s20_t_120.npy
│
└───models  # pretrained network weights for the occupancy-based prior
        people_net_w64_s8_t_120.pth
        people_net_w64_s16_t_120.pth
        people_net_w64_s20_t_120.pth

In our experiments we used the following two datasets:

• ATC pedestrian tracking dataset (ATC): this dataset comprises of 92 days in total. For our experiments, we only used data from Wednesday November 14th, 2012 (training) and Saturday November 18th, 2012 (testing). Moreover, we curated the data for ease of processing, downsampling the amount of measurements by a factor of 5. We provide our curated datasets in the data/ATC/subsampled folder. The occupancy map for ATC is found in the data/ATC folder.
• KTH track dataset (KTH): we curated this dataset to make it consistent with the ATC format. We provide our curated datasets in the data/KTH_track folder. The occupancy map for KTH is found in the same folder.

Moreover, in the models folder, we provide the pretrained network models for the occupancy-based prior, used in the experiments. The filenames for these models follow the structure people_net_w{window_size}_s{scale_factor}_t_{number_of_training_epochs}.pth

Finally, in the maps folder, we provide the precomputed occupancy-based prior MoDs obtained by using each of the deep models on each environment. The filenames for these maps follow the structure map_{dataset}_w{window_size}_s{scale_factor}_t_{number_of_training_epochs}.npy

For both models and maps, the scale_factor is indicated as multiplier from the occupancy map scale of 0.05m/pixel, so:

• s8 => 8x0.05 = 0.4m/cell corresponds to the prior $\bar{\textbf{d}}_{0.4}^W$ from the paper,
• s16 => 16x0.05 = 0.8m/cell to $\bar{\textbf{d}}_{0.8}^W$, and
• s20 => 20x0.05 = 1.0m/cell to $\bar{\textbf{d}}_{1.0}^W$.

Replicating the experiments

Once you have downloaded the data and ensure paths are correct, you can use the scripts in this codebase to replicate our experiments, following these steps:

1. Run build_bayesian_discrete_directional_atc.py and build_bayesian_discrete_directional_kth.py to generate the trajectory-based floor field models from the curated data.
2. (optional) Run training_atc.py to train your own occupancy-based prior network. This step can be skipped by using the provided pretrained models.
3. (optional) Run map_occupancy_prior_atc.py and map_occupancy_prior_kth.py to build the MoDs for the occupancy-based prior. This step can be skipped by using the provided precomputed maps.
4. Run compute_curves_atc.py and compute_curves_kth.py to compute the likelihood performance of BFF compared to the trajectory-based floor field model without occupancy-based prior.
5. Run plot_curves.py to obtain the plots presented in Fig. 3 in the paper.
6. Run plot_quivers_atc.py and plot_quivers_kth.py to obtain the plots presented in Fig. 4 in the paper.

Citation

If you find this work useful, please consider citing:

@inproceedings{202410_verdoja_bayesian,
  title       = {{Bayesian} {Floor} {Field}: {Transferring} people flow predictions across environments},
  booktitle   = {2024 {IEEE}/{RSJ} {International} {Conference} on {Intelligent} {Robots} and {Systems} ({IROS})},
  author      = {Verdoja, Francesco and Kucner, Tomasz Piotr and Kyrki, Ville},
  month       = oct,
  year        = {2024},
}

Moreover, if you use OctoPytorch, or either dataset mentioned above, please consider citing the original authors as well.

List of citations

@article{brvsvcic_2013_atc,
  title={Person tracking in large public spaces using 3-D range sensors},
  author={Br{\v{s}}{\v{c}}i{\'c}, Dra{\v{z}}en and Kanda, Takayuki and Ikeda, Tetsushi and Miyashita, Takahiro},
  journal={IEEE Transactions on Human-Machine Systems},
  volume={43},
  number={6},
  pages={522--534},
  year={2013},
  publisher={IEEE}
}

@inproceedings{dondrup_2015_kth,
  title = {Real-Time Multisensor People Tracking for Human-Robot Spatial Interaction},
  author = {Dondrup, Christian and Bellotto, Nicola and Jovan, Ferdian and Hanheide, Marc},
  booktitle = {International Conference on Robotics and Automation (ICRA) - Workshop on Machine Learning for Social Robotics},
  year = {2015}
}

@inproceedings{Jegou_2017_tiramisu,
  title = {The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation}, 
  author = {Jégou, Simon and Drozdzal, Michal and Vazquez, David and Romero, Adriana and Bengio, Yoshua},
  booktitle = {2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)}, 
  year = {2017},
  pages = {1175-1183},
  doi = {10.1109/CVPRW.2017.156}
}

Acknowledgements

This work is part of the Hypermaps project and was supported by the Research Council of Finland, decision 354909.
We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.