https://github.com/vita-epfl/trajnetplusplusbaselines

[ITS'21] Human Trajectory Forecasting in Crowds: A Deep Learning Perspective
https://github.com/vita-epfl/trajnetplusplusbaselines

benchmark crowd-dynamics datasets human-trajectory human-trajectory-prediction kalman-filter orca social-force-model social-gan social-lstm trajectory-prediction-benchmark

Last synced: 5 months ago
JSON representation

[ITS'21] Human Trajectory Forecasting in Crowds: A Deep Learning Perspective

• Host: GitHub
• URL: https://github.com/vita-epfl/trajnetplusplusbaselines
• Owner: vita-epfl
• License: mit
• Created: 2019-10-28T09:42:36.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2023-04-04T15:30:41.000Z (about 3 years ago)
• Last Synced: 2023-11-07T16:56:49.249Z (over 2 years ago)
• Topics: benchmark, crowd-dynamics, datasets, human-trajectory, human-trajectory-prediction, kalman-filter, orca, social-force-model, social-gan, social-lstm, trajectory-prediction-benchmark
• Language: Python
• Homepage: https://ieeexplore.ieee.org/document/9408398
• Size: 41.3 MB
• Stars: 216
• Watchers: 14
• Forks: 79
• Open Issues: 7
• Metadata Files:
  • Readme: README.rst
  • License: LICENSE

Awesome Lists containing this project

README

          
TrajNet++ : The Trajectory Forecasting Framework

================================================

PyTorch implementation of `Human Trajectory Forecasting in Crowds: A Deep Learning Perspective `_ 

.. figure:: docs/train/cover.png

TrajNet++ is a large scale interaction-centric trajectory forecasting benchmark comprising explicit agent-agent scenarios. Our framework provides proper indexing of trajectories by defining a hierarchy of trajectory categorization. In addition, we provide an extensive evaluation system to test the gathered methods for a fair comparison. In our evaluation, we go beyond the standard distance-based metrics and introduce novel metrics that measure the capability of a model to emulate pedestrian behavior in crowds. Finally, we provide code implementations of > 15 popular human trajectory forecasting baselines.

We host the `Trajnet++ Challenge `_ on AICrowd allowing researchers to objectively evaluate and benchmark trajectory forecasting models on interaction-centric data. We rely on the spirit of crowdsourcing and the challenge has > 1800 submissions. We encourage researchers to submit their sequences to TrajNet++, so the quality of trajectory forecasting models can keep increasing in tackling more challenging scenarios.  

Data Setup

==========

The detailed step-by-step procedure for setting up the TrajNet++ framework can be found `here `_

Converting External Datasets

----------------------------

To convert external datasets into the TrajNet++ framework, refer to this `guide `_ 

Training Models

===============

LSTM

----

The training script and its help menu:

``python -m trajnetbaselines.lstm.trainer --help``

**Run Example**

.. code-block::

   ## Our Proposed D-LSTM

   python -m trajnetbaselines.lstm.trainer --type directional --augment

   ## Social LSTM 

   python -m trajnetbaselines.lstm.trainer --type social --augment --n 16 --embedding_arch two_layer --layer_dims 1024

GAN

---

The training script and its help menu:

``python -m trajnetbaselines.sgan.trainer --help``

**Run Example**

.. code-block::

   ## Social GAN (L2 Loss + Adversarial Loss)

   python -m trajnetbaselines.sgan.trainer --type directional --augment

   

   ## Social GAN (Variety Loss only)

   python -m trajnetbaselines.sgan.trainer --type directional --augment --d_steps 0 --k 3

Evaluation

==========

The evaluation script and its help menu: ``python -m evaluator.lstm.trajnet_evaluator --help``

**Run Example**

.. code-block::

   

   ## LSTM (saves model predictions. Useful for submission to TrajNet++ benchmark)

   python -m evaluator.lstm.trajnet_evaluator --output OUTPUT_BLOCK/trajdata/lstm_directional_None.pkl --path 

   ## SGAN (saves model predictions. Useful for submission to TrajNet++ benchmark)

   python -m evaluator.sgan.trajnet_evaluator --output OUTPUT_BLOCK/trajdata/sgan_directional_None.pkl --path 

More details regarding TrajNet++ evaluator are provided `here `_

Evaluation on datasplits is based on the following `categorization `_

Results

-------

Unimodal Comparison of interaction encoder designs on interacting trajectories of TrajNet++ real world dataset. Errors reported are ADE / FDE in meters, collisions in mean % (std. dev. %) across 5 independent runs. Our goal is to reduce collisions in model predictions without compromising distance-based metrics.

+----------------+------------+-------------------+ 

| Method         |   ADE/FDE  | Collisions        | 

+----------------+------------+-------------------+ 

| LSTM           |  0.60/1.30 | 13.6 (0.2)        | 

+----------------+------------+-------------------+ 

| S-LSTM         |  0.53/1.14 |  6.7 (0.2)        |  

+----------------+------------+-------------------+ 

| S-Attn         |  0.56/1.21 |  9.0 (0.3)        |  

+----------------+------------+-------------------+ 

| S-GAN          |  0.64/1.40 |  6.9 (0.5)        |   

+----------------+------------+-------------------+ 

| D-LSTM (ours)  |  0.56/1.22 |  **5.4** **(0.3)**| 

+----------------+------------+-------------------+ 

Interpreting Forecasting Models

===============================

+-------------------------------------------------------------------------+

|  .. figure:: docs/train/LRP.gif                                         |

|                                                                         |

|     Visualizations of the decision-making of social interaction modules |

|     using layer-wise relevance propagation (LRP). The darker the yellow |

|     circles, the more is the weight provided by the primary pedestrian  |

|     (blue) to the corresponding neighbour (yellow).                     |

+-------------------------------------------------------------------------+

Code implementation for explaining trajectory forecasting models using LRP can be found `here `_

Benchmarking Models

===================

We host the `Trajnet++ Challenge `_ on AICrowd allowing researchers to objectively evaluate and benchmark trajectory forecasting models on interaction-centric data. We rely on the spirit of crowdsourcing, and encourage researchers to submit their sequences to our benchmark, so the quality of trajectory forecasting models can keep increasing in tackling more challenging scenarios.

Citation

========

If you find this code useful in your research then please cite

.. code-block::

    

    @article{Kothari2020HumanTF,

      author={Kothari, Parth and Kreiss, Sven and Alahi, Alexandre},

      journal={IEEE Transactions on Intelligent Transportation Systems}, 

      title={Human Trajectory Forecasting in Crowds: A Deep Learning Perspective}, 

      year={2021},

      volume={},

      number={},

      pages={1-15},

      doi={10.1109/TITS.2021.3069362}

     }