TexSense - Running

The project aims to predict continuous 3D knee moments during various running conditions using pressure insoles (PI) and inertial measurement unit (IMU) data. This repo includes implementations of the neural network models, hyperparameter optimization, data preprocessing utilities, and evaluation metrics for assessing model performance. The goal is to develop accurate models that can assist in biomechanical analysis and injury prevention in sports and rehabilitation contexts.

The project TexSense is funded within the context of WISS 2025, der Wissenschafts- und Innovationsstrategie 2025, by the federal state of Salzburg.

Publications:

• Höschler, L., Halmich, C., Schranz, C., Fritz, J., Koelewijn, AD., Schwameder, H. (2024). TOWARDS REAL-TIME ASSESMENT: WEARABLE-BASED ESTIMATION OF 3D KNEE KINETICS IN RUNNING AND THE INFLUENCE OF PREPROCESSING WORKFLOWS. ISBS Proceedings Archive: Vol. 37: Iss.2, ISBS 2024 Conference in Salzburg, 15th - 20th July, 2024. https://commons.nmu.edu/isbs/vol42/iss1/72/
• Höschler, L., Halmich, C., Schranz, C., Fritz, J., Čigoja, S., Ullrich, M., Koelewijn, AD., Schwameder, H. (2025). Wearable-based estimation of continuous 3D knee moments during running using a convolutional neural network. Sports Biomechanics, 1-19. https://doi.org/10.1080/14763141.2025.2481164
• Höschler, L., Halmich, C., Schranz, C., Koelewijn, AD., Schwameder, H. (2025). Evaluating the Influence of Sensor Configuration and Hyperparameter Optimization on Wearable-Based Knee Moment Estimation During Running. International Journal of Computer Science in Sport, 24 (2). 80-106. https://doi.org/10.2478/ijcss-2025-0014

Data Set

The repo comes with an example dataset containing a subset of 10 participants of our original data (n=19). Only a single shoe condition is provided. The data consists of multiple 10 second - samples of continuous (synchronized and preprocessed) IMU, PI and 3D Knee moments data from each running condition. Detailed information regarding preprocessing are stated below.

DOI: 10.5281/zenodo.11119845
Link to example data: example dataset

Pre-Processing Information

3D Knee Moments:

• 3D Inverse Dynamics during stance phases
• flight phases filled with zero-values
• fs = 100 Hz (downsampled from 200 Hz)
• normalized to body mass
• tensors with shape [1000 x 3]

PI Data

• "raw" pressure readings from XSensor sensors (resolution: 31 x 11)
• fs = 100 Hz
• normalized to body mass
• flattened tensors with shape [1000 x 341]

IMU Data

• 3D Accelerometer and Gyroscope readings from 4 body positions (foot, shank, thigh, pelvis)
• filtered using a 2nd order Butterworth low-pass filter with custom cut-off frequencies
• each channel is z-score normalized
• tensors with shape [4000 x 24]

Conditions

Treadmill Slopes:

• 0 % (level)
• 5 % incline (up)
• 5 % decline (down)
  

Treadmill Speeds:

• self-selected speed
• self-selected speed - 1 km/h
• self-selected speed + 1 km/h
  

Sides:

• left
• right
  

Slices:

• non-overlapping 10-second windows over 60 seconds of running
  

Naming convention: ParticipantID_Shoe_Slope_Speed_Side_Slice

Table of Contents

1. Installation
2. Usage
3. File Structure
4. Contributing
5. License

Installation

1. Clone the repository:

   git clone https://github.com/luceskywalker/TexSense.git

2. Install the required dependencies:

   pip install -r requirements.txt

Usage

Data Preparation

1. Ensure you have downloaded the example dataset into the CONT/ directory.

Model Training and Evaluation

1. Train a model using the provided script or notebook files in the main/ directory.
2. Select specific sensor configurations by editing the [sensor_setup] section in config/CONT_config.txt
3. Evaluate the trained model using the model_statistics() function in model_stats.py.
4. Plot the training and validation loss using plot_loss() function in plot.py
5. Plot the predictions loss using plot_prediction() function in plot.py

Hyperparameter Optimization

1. Conduct hyperparameter optimization experiments using the main/main_hyperopt.py script
2. Specify search spaces and optimization params in the script
3. Select specific sensor configurations by editing the [sensor_setup] section in config/CONT_config.txt
4. Logfiles are created in the hyperopt/log/ directory.

Predict with Pre-Trained Models

1. Load one of the pre-trained models provided in model/ and predict knee moments from IMU data
2. Select specific subjects for validation by altering the VAL_SUBJ parameter
3. Evaluate model performance over continuous predictions by RMSE and nRMSE using the model_statistics() function in model_stats.py.
4. Plot the predictions loss using plot_prediction() function in plot.py
5. Note: predictions may look "too good" because the subjects might have been included in the training set used for training the models.

Run own experiments

1. Alter the model architecture, sensor configuration, or training parameters by editing the CONT_config.txt in config/

File Structure

projectname/
│
├── CONT/
│   ├── config/ 
│   │   └── CONT_config.txt    # Configuration File
│   │
│   ├── hyperopt/ 
│   │   └── log/               # Location where hyperopt log files are created
│   │
│   ├── main/
│   │   ├── main_CONT.ipynb    # main as .ipynb
│   │   ├── main_CONT.py       # main as .py
│   │   ├── main_hyperopt.py   # hyperopt main script 
│   │   └── predict_from_pretrained_models.py  # use pretrained models to predict
│   │
│   ├── utils/
│   │   ├── data.py            # dataset and dataloader
│   │   ├── loss.py            # custom loss function
│   │   ├── model.py           # cnn model
│   │   ├── plot.py            # plot functions
│   │   ├── training.py        # train and eval routines
│   │   └── utils.py           # preprocessing and evaluation utils
│   │
│   ├── models/
│   │   ├── foot.pt            # pretrained model using only foot imu
│   │   ├── foot_shank.pt      # pretrained model using only foot and shank imu
│   │   ├── foot_shank_pelvis.pt  # pretrained model using only foot, shank and pelvis imu
│   │   └── foot_shank_thigh_pelvis.pt  # pretrained model using only foot, shank, thigh and pelvis imu
│   │
│   └── example_dataset/       # Raw dataset files
│
├── README.md                  # Project README file
├── requirements.txt           # Project dependencies
├── LICENSE_code.md            # License information (code)
└── LICENSE_data.md            # License information (data)

Contributing

Contributions to the project are welcome! Please reach out to us via e-mail.

License

Please note that the TexSense repository is dual-licensed.
The TexSense dataset is available under the terms of the Creative Commons BY-NC 4.0 license.
The provided example code is licensed under the MIT license.

Please see the respective LICENSE files for more details.

Contact

Lucas Höschler, lucas.hoeschler@googlemail.com