🚀 MCST: Maneuver Compensation Strong Tracker

MCST is an adaptive deep learning-based radar target tracking algorithm designed to handle high-speed and highly maneuverable targets. It leverages a Bi-LSTM architecture with a dual-level attention mechanism and a custom Maneuver Compensation Unit (MCU) to achieve robust and accurate tracking performance, even under extreme target dynamics and observation uncertainties.

📄 Accepted at IEEE Transactions on Aerospace and Electronic Systems
📚 Read the paper

✨ Features

🚀 Supports tracking of hypersonic, highly maneuverable targets
🧠 Built upon a Bi-LSTM architecture with:
- Predictor & Updater modules inspired by Kalman filtering
- Dual-level attention module for temporal and sample weighting
- Maneuver Compensation Unit (MCU) using FFT-based residual analysis
🔄 Outputs both state estimates and uncertainty covariance matrices
📈 Trained and tested on a custom 10,000-trajectory dataset
📦 Modular and extensible codebase

🖼️ Model Architecture

MCST Architecture
A hybrid Bi-LSTM + attention framework with maneuver compensation and uncertainty modeling.

📁 Project Structure

MCST/
├── config.py            # Model and training configs
├── data/                # Dataset and preprocessing scripts
├── models/              # Model definitions (Predictor, MCU, Updater)
├── utils/               # Utilities (normalization, trajectory initialization)
├── train/               # Training pipeline
├── evaluation/          # Evaluation and metrics
├── log/                 # log
├── main.py

🧪 Quick Start

1. Setup Environment

conda create -n mcst python=3.9
conda activate mcst

2. Download Dataset

You can download from https://github.com/Shen-kl/OneManeuveringTarget3D.

3. Train the Model

python main.py

4. Evaluate

python evaluate.py --checkpoint checkpoints/mcst_best.pth

📊 Dataset Overview

MCST is trained on a synthetic dataset of 10,000 trajectories, simulating targets with:

Velocities up to Mach 5
Acceleration range: 3g–7g
Motion models: CV, CA, HCT, FCT
Sampling interval: 0.4 seconds
Noise added in spherical coordinates, then converted to Cartesian

Each trajectory contains 100 frames, with 3 model transitions per trajectory.

📈 Performance

MCST outperforms several SOTA model-based tracking algorithms, including:

Single model: FSTCKF ,and RNSTF
Multiple models: HGMM, and RIMM

Especially in miss detection and rapid maneuver scenarios.

📜 Citation

If you find this work helpful, please cite our paper:

@article{shen2024mcst,
  title={MCST: An adaptive tracking algorithm for high-speed and highly maneuverable targets based on bidirectional LSTM network},
  author={Shen, Kailun and Yuan, Weiming and Yan, Junkun and Ma, Keke},
  journal={IEEE Transactions on Aerospace and Electronic Systems},
  year={2024},
  publisher={IEEE}
}

📜 Changelog

2025/9/18

Added

We have added the functionality to export the model in TorchScript format.

Changed

The computation of the location-dependent min–max bounds in min–max normalization has been revised. Specifically, the original scheme that relied on a fixed maximum velocity has been replaced by a hybrid strategy that blends the fixed maximum velocity with the model-estimated filtered velocity.

Fixed

Fixed the lack of normalization for the gain matrix $W$ in the Update class. This yields more balanced gradient distributions across all model layers.

MCST

Install / Use

README