Trajectory Prediction for Heterogeneous Agents: A Performance Analysis and on Small and Imbalanced Datasets

Paper

Agent classes are powerful cues for trajectory prediction.

Project Organization

├── LICENSE
├── Makefile           <- Makefile with commands like `make data` or `make train`
├── README.md          <- The top-level README for developers using this project.
├── data
│   ├── original       <- Data to train the models
│   ├── processed      <- Data to evaluate k-fold cross validation
│
├── notebooks          <- Jupyter notebooks. 
│
├── requirements.yml   <- The requirements file for reproducing the analysis environment
│                         
├── setup.py           <- makes project pip installable (pip install -e .) so src can be imported
├── class_cond_trajpred                <- Source code for use in this project.
│   ├── __init__.py    <- Makes src a Python module
│   │
│   ├── models         <- Scripts to train models and then use trained models to make
│   │   │                 predictions
│   │   └── train_model.py
│   ├── mods         <- Scripts to train and evaluate MoD-based models
│   │   │                 
│   │   └── train_model.py
│   │
│   └── visualization  <- Scripts to create exploratory and results oriented visualizations
│       └── visualize.py
└── tox.ini            <- tox file with settings for running tox; see tox.readthedocs.io

---

Project based on the cookiecutter data science project template. #cookiecutterdatascience

Install

Install miniconda. Then, you can install all packages required by running:

conda env create -f requirements.yml && conda activate class-cond-trajpred

And run:

pip install .

Training and testing deep learning predictors

Steps to train and test deep learning-based predictors (i.e., RED, cRED, TF, cTF, GAN, cGAN, VAE, and cVAE).

Train one model on a dataset

Set the corresponding cfg file on the config folder to the original data as follows:

---

data :
  data_dir : data/original/thor_magni/  # data/original/sdd/ for SDD dataset

---

Run the train_model module with the corresponding config file:

python -m class_cond_trajpred.data_modeling.train_model class_cond_trajpred/cfg/thor_magni/sup_ctf.yaml 

k-fold cross validation

Set the corresponding cfg file on the config folder to the processed data as follows:

---

data :
  data_dir : data/processed/sdd/k_fold_10  # data/original/thor_magni/ for THOR-MAGNI dataset

---

python -m class_cond_trajpred.data_modeling.k_fold_cv 10 class_cond_trajpred/cfg/thor_magni/gan.yaml

To train with (x, y, v_x, v_y), we need to change the cfg file under network: set observation_len to 8.

Training and testing MoD-based predictors

In MoD-based predictors, we have class-conditioned CLiFF-LHMP and general CLiFF-LHMP. Both methods use same evaluation datasets: SDD and THÖR-MAGNI. The corresponding dataset directories are located at data/mod_based_predictors_data/sdd and data/mod_based_predictors_data/thor_magni. Within the dataset folder, test ratios ranging from 10% to 90% are available. For instance, k_fold_10_percentage_test corresponds to configurations where 10% of the data is used for testing and the remaining 90% for training. In each test ratio configuration, the test data are repeatedly and randomly sub-sampled 10 times, resulting in batches from batch1 to batch10.

To run MoD predictors:

./mod_based_predictors/run_mod_predictors.sh

In run_mod_predictors.sh, it contains command to run MoD predictors for both datasets, take class-conditioned CLiFF-LHMP for THÖR-MAGNI dataset as an example:

r_s=0.2
exp_type="condition" 
if_rank=2
generate_traj_num=5
beta=1

version="r${r_s}_k${generate_traj_num}_${exp_type}"
python3 mod_based_predictors/main_mod_magni.py "$r_s" "$generate_traj_num" "$version" "$if_rank" "$exp_type" "$beta"

Here, the input parameters are:

• r_s: sampling radius, for sampling velocity in MoD

• exp_type: "condition" or "general"

• if_rank:

  • 0: for evaluate with k predicted trajectory, output average ADE/FDE
  • 1: for evaluate with k predicted trajectory, output weighted average ADE/FDE
  • 2: for evaluate with most-likely predicted trajectory,

• generate_traj_num: generate k predicted trajectories

• beta: for controlling the reliance on the MoD versus the CVM,with a lower β favoring the velocity sampled from the MoD.

The output evaluation metrics are saved in mod_based_predictors/results. It is also optional to save and plot predicted trajectories.