Deep Learning 1D-CNN-Based Ground Contact Detection in Sprint Acceleration Using Inertial Measurement Units

Area of Science:

• Biomechanics
• Sports Science
• Machine Learning

Background:

• Accurate ground contact (GC) detection is crucial for sprint performance analysis.
• Inertial measurement units (IMUs) offer field-based assessment but have limitations in sprint acceleration reliability.
• Existing heuristic and machine learning algorithms struggle with precise GC detection during this phase.

Purpose of the Study:

• To introduce and evaluate a deep learning one-dimensional convolutional neural network (1D-CNN) for enhanced GC event and time detection.
• To improve the reliability of IMU-based biomechanical monitoring during sprint acceleration.
• To compare the 1D-CNN's performance against existing methods.

Main Methods:

• Twelve athletes performed 60 m sprints, with the initial 15 m captured by bilateral shank-mounted IMUs and high-speed video.
• Video-derived GC events were used as ground truth for training and validating the 1D-CNN model.
• Resultant acceleration and angular velocity from IMUs served as input features for the deep learning model.

Main Results:

• The optimized 1D-CNN model demonstrated excellent performance with mean Hausdorff distances ≤ 6 ms.
• The model achieved 100% precision and recall, with a Rand Index ≥ 0.977 on validation and test datasets.
• Near-perfect agreement with video references was observed (bias < 1 ms, limits of agreement ± 15 ms, r > 0.90).

Conclusions:

• The 1D-CNN significantly outperforms heuristic and previous machine learning methods for GC detection in sprint acceleration.
• This deep learning approach provides robust and highly accurate GC detection, suitable for real-world biomechanical monitoring.
• The findings underscore the potential of deep learning time-series models for advancing sprint analysis and performance optimization.