Multi-Ship Trajectory Prediction Based on Social-LSTM

Project details

基于Social-LSTM的多船舶协同轨迹预测模型实现，本项目使用PyTorch构建了Social-LSTM模型，进行实验验证了考虑船舶间相互影响的Social-LSTM轨迹预测模型相比于普通LSTM模型的预测准确性。

PyTorch implementation for Social-LSTM, which is built to predict multi-vessel trajectories. Experiments have been done to demonstrate that Social-LSTM can predict better trajectories than LSTM.

This project has been forked initially from https://github.com/quancore/social-lstm. If you find this code useful in your research, please cite the paper CVPR16_Social_LSTM.

Documentation

1. criterion.py: Python script for loss functions.
   Including a GaussianLikehood loss function and a RMSE loss function.
2. utils.py: Python script for handling input train/test/validation data and preprocess it.
   DataLoader class includes time_preprocess function, data load function, batch function and other data process function.
3. model.py: Python script includes Social-LSTM and Vanilla-LSTM.
   Social-LSTM model implementation, Vanilla-LSTM model implementation and related functions.
4. helper.py: Python script includes various helper methods.
5. train.py: Python script for training Social-LSTM model.
6. train_vlstm.py: Python script for training Vanilla-LSTM model.
7. test.py: Python script for model testing and getting output txt file for submission.
8. validation.py: Python script for externally evaluate a trained model by getting validation error.
9. visualize.py: Python script for visualizing predicted trajectories during train/test/validation sessions.

Results

Model	Neighbor Size	Mean Error	Final Error
Vanilla-LSTM	0	0.6430	2.0371
Social-LSTM	0.021	0.6323	2.0572
Social-LSTM	0.020	0.6363	1.9084
Social-LSTM	0.019	0.6422	2.0148

Subjective display of predicted trajectories