DSPIGCN: Dual-stream Physics-informed Graph Convolutional Network for Reliable Pedestrian Trajectory Prediction

DSPIGCN: Dual-stream Physics-informed Graph Convolutional Network for Reliable Pedestrian Trajectory Prediction Introduction DSPIGCN is a deep learning model for pedestrian trajectory prediction that incorporates physics-informed constraints. Setup Prerequisites Python 3.8 PyTorch CUDA (if using GPU) Installation Install the required dependencies: pip install -r requirements.txt Data Preparation The project expects the datasets to be located in the datasets/ directory. The supported datasets include: eth hotel univ zara1 zara2 trajnet sdd Ensure your data is organized as follows: datasets/ eth/ test/ train/ val/ ... Training To train the model, use train.py. You can specify various hyperparameters and the target dataset. Common Arguments --dataset: Dataset to train on (e.g., eth, hotel, univ, zara1, zara2). --tag: A unique tag for the model checkpoint directory. --num_epochs: Number of training epochs. Default: 200. --batch_size: Batch size. Default: 128. --lr: Learning rate. Default: 0.001. --use_lrschd: Enable learning rate scheduler. --n_stgcnn: Number of ST-GCNN layers. Default: 1. --n_txpcnn: Number of TXPCNN layers. Default: 5. --obs_seq_len: Observation sequence length. Default: 8. --pred_seq_len: Prediction sequence length. Default: 12. Example Commands You can refer to train.sh for batch training commands. Here are examples for individual datasets: Train on ETH: python train.py --lr 0.001 --n_stgcnn 1 --n_txpcnn 5 --dataset eth --tag social-stgcnn-eth --use_lrschd --num_epochs 200 Checkpoints will be saved in KDD_checkpoint/<tag>/. Evaluation To evaluate the trained models, use test.py. python test.py Results Evaluation results (ADE, FDE, etc.) will be saved in the Results/KDD/ directory.