xLength: Predicting Expected Ski Jump Length Shortly after Take-Off Using Deep Learning

Johannes Link, Leo Schwinn, Falk Pulsmeyer, Thomas Kautz, Bjoern M. Eskofier

TL;DR

This paper introduces "xLength," a novel deep learning model that predicts ski jump length immediately after take-off using wearable sensor data. By leveraging this technology, the authors demonstrate the feasibility of real-time performance prediction, offering valuable insights for live broadcasting and expert analysis in ski jumping.

What Problem Does This Paper Address?

Ski jumping, a sport characterized by high speeds and complex aerodynamics, has historically relied on manual analysis and limited tracking data for performance evaluation. While sports analytics has seen a surge in data-driven approaches, ski jumping remains relatively understudied in this regard. The challenge lies in accurately and efficiently predicting a crucial performance metric – jump length – in real-time, shortly after the critical take-off phase. Existing tracking systems are often expensive, weather-dependent, or lack the necessary precision for such a dynamic sport. This paper aims to bridge this gap by developing a deep learning model that can predict jump length using unobtrusive wearable sensors, providing a more accessible and data-rich approach to performance analysis.

Methods

The authors explore four deep learning architectures: a Fully Connected Network (FCN), two Convolutional Neural Networks (CNNs), and a Long Short-Term Memory (LSTM) network. The FCN serves as a baseline, while the CNNs and LSTM are investigated for their ability to leverage temporal information from sensor data. The input to these models includes 3D position, 3D velocity, and ski orientation data from wearable trackers, augmented with metadata such as gender, wind conditions, and venue-specific parameters.

A key aspect of their methodology is the rigorous hyperparameter tuning using Bayesian optimization within a nested 5-fold cross-validation framework. This approach is crucial for preventing overfitting and obtaining reliable performance estimates, especially when generalizing to unseen athletes and venues. Data augmentation techniques, such as mirroring and Gaussian noise, are applied cautiously due to the sensitive nature of ski jump dynamics. Standardization of data is also employed.

Datasets

The study utilizes a newly compiled dataset of 2523 ski jumps from 205 professional athletes across five different ski jumping venues. This dataset is unique in its scale and the use of wearable sensors (IMUs and ultra-wideband technology) to capture detailed 3D trajectory, velocity, and ski orientation data. This dataset is referred to as the "xLength dataset" and serves as both the training and evaluation benchmark. The authors note that while this dataset is larger than previous ones in ski jumping tracking, the number of venues is relatively small and unevenly distributed.

Key Insights and Ideas

  • Real-time Prediction is Feasible: The core idea is that predicting jump length shortly after take-off is achievable with current sensor technology and deep learning, opening doors for real-time applications.
  • ResNet Shows Strong Generalization: The ResNet architecture consistently demonstrated the best performance, particularly in generalizing to new athletes, suggesting its robustness in capturing underlying jump dynamics.
  • Input Length Matters, But with Diminishing Returns: While longer input sequences (more time after take-off) generally improve prediction accuracy, the gains become less significant after approximately 3 seconds, indicating a point where additional data offers less new information for predicting the final jump length.
  • Venue Generalization is Challenging but Achievable: Predicting jump length for unseen venues is more difficult than for unseen athletes, likely due to the uneven distribution of data across venues. However, the performance is still promising, suggesting that with a limited number of professional venues, generalization is practical.
  • Wearable Sensors Offer Unobtrusive Data: The use of wearable trackers on skis is highlighted as an unobtrusive method, ideal for competition settings without interfering with athlete performance.

Key Results

  • Athlete Generalization: The ResNet achieved a Mean Absolute Error (MAE) of 5.3 m when generalizing to unseen athletes, with a mean residual of 0.1 m and a standard deviation of 6.8 m.
  • Venue Generalization: For generalization to unseen venues, the ResNet achieved the lowest MAE of 5.9 m, with a mean residual of 0.7 m and a standard deviation of 7.6 m. The LSTM struggled significantly with venue generalization, showing a much higher MAE of 9.2 m.
  • Input Length Impact: The standard deviation of the prediction error decreased from 0.5 s to 4.0 s after take-off, indicating improved prediction consistency with longer input sequences. The mean prediction error remained relatively constant.
  • Dataset Characteristics: The dataset exhibits two peaks in jump length distribution around 93 m and 120 m, with a median jump length of 97.3 m.

Limitations and Future Work

  • Limited Venue Data: The dataset, while large, has a relatively small number of venues, which impacts venue generalization. Future work could involve acquiring data from more venues or utilizing transfer learning from other sports.
  • Low Sampling Rate: The 20 Hz sampling rate of the UWB measurements is considered relatively low for capturing the high speeds at take-off. Accessing the higher 1000 Hz IMU data could provide more detailed insights into the take-off phase.
  • Body Movement Ignored: The current sensor placement on skis does not capture the athlete's body movement, which is crucial for take-off analysis. Future work could incorporate camera-based pose estimation or additional body-worn sensors.
  • Explainability: While the models achieve good prediction accuracy, interpreting their decisions remains a challenge. Future research could focus on more explainable AI approaches (e.g., SHAP) to understand feature importance.
  • Ski Flying Extension: The authors suggest extending the dataset and prediction to ski flying, which presents its own data acquisition challenges.
  • Live Broadcasting Integration: The paper highlights the potential for live broadcasting and expert analysis, suggesting further development in integrating xLength into these applications.

Summary updated

Abstract

With tracking systems becoming more widespread in sports research and regular training and competitions, more data are available for sports analytics and performance prediction. We analyzed 2523 ski jumps from 205 athletes on five venues. For every jump, the dataset includes the 3D trajectory, 3D velocity, skis’ orientation, and metadata such as wind, starting gate, and ski jumping hill data. Using this dataset, we aimed to predict the expected jump length (xLength) inspired by the expected goals metric in soccer (xG). We evaluate the performance of a fully connected neural network, a convolutional neural network (CNN), a long short-term memory (LSTM), and a ResNet architecture to estimate the xLength. For the prediction of the jump length one second after take-off, we achieve a mean absolute error (MAE) of 5.3 m for the generalization to new athletes and an MAE of 5.9 m for the generalization to new ski jumping hills using ResNet architectures. Additionally, we investigated the influence of the input time after the take-off on the predictions’ accuracy. As expected, the MAE becomes smaller with longer inputs. Due to the real-time transmission of the sensor’s data, xLength can be updated during the flight phase and used in live TV broadcasting. xLength could also be used as an analysis tool for experts to quantify the quality of the take-off and flight phases.
BibTeX
@inproceedings{link2022xlength,
  author = {Johannes Link and Leo Schwinn and Falk Pulsmeyer and Thomas Kautz and Bjoern M. Eskofier},
  title = {{xLength: Predicting Expected Ski Jump Length Shortly after Take-Off Using Deep Learning}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=rL6ID6rqV0},
}

Loading related content...

Figure