DIMM

This repo is the source code and supplementary material of the paper "DIMM: Decoupled Multi-hierarchy Kalman Filter via Reinforcement Learning".

Prerequisites

Before you begin, ensure you have met the following requirements:

• Python: This project is primarily developed in Python. Make sure you have Python installed on your system. You can download it from python.org.
• Python Libraries: It is recommended to build an environment either in pip or conda and install the following packages:
  $\quad$ - Tensorflow (2.17.0)
  $\quad$ - Gym (0.26.2)
  $\quad$ - Matplotlib (3.8.4)
  $\quad$ - Numpy (1.26.4)
  

Directory structure

DIMM/ # Decoupled IMM (interacting multiple model)
│
├── Filters/
│   ├── KF.py      # basic Kalman filtering algorithm
│   └── ourIMM.py  # IMM algorithm with CV, CA, and CJ model
│
├── data/ 
│   ├── OKFdata/   # OKF dataset
│   │   ├── train  # train set
│   │   ├── test   # test set
│   └── └── val    # validation set
│
├── RLEnvs/
│   └──Estimation.py  # RL environment: position estimation
│
├── RL/
│   └──TD3.py  # RL method: attention-based twin delayed deep deterministic policy gradient (TD3)
│
├── models/  # save the best training model
│
├── train.py    # main function for training the 'DIMM' model
├── test.py     # main function for testing the 'DIMM' model
└── Metrics.py  # evaluation metrics to test the algorithm's estimation accuracy

Running

1. Run 'train.py' to train the best model of the actor and critic networks in DIMM and save them in the 'models' folder.
2. Run 'test.py' to test the trained DIMM model, compare its estimates with the ground-truth values, and calculate its estimation errors.

Citation

If you use this project in your research, please cite the following paper:

@article{zha2025dimm,
  title={DIMM: Decoupled Multi-hierarchy Kalman Filter for 3D Object Tracking},
  author={Zha, Jirong and Fan, Yuxuan and Li, Kai and Li, Han and Gao, Chen and Chen, Xinlei and Li, Yong},
  journal={arXiv preprint arXiv:2505.12340},
  year={2025}
}