Text

Introduction: Musculoskeletal injuries represent one of the leading causes of dropout from military training and operational duty, resulting in significant financial costs and reduced operational readiness for military organisations. Detailed monitoring of physical load during training and operational tasks has the potential to reduce this musculoskeletal injury burden. Accelerometry can objectively assess human physical activity by providing estimates of an activity’s duration, frequency, and intensity. However, contextual information about the acquired data is necessary. Therefore, this study aims to classify the most relevant military police activities using a single body-fixed tri-axial accelerometer in a practical setting.

Methods: Twenty-three participants executed a protocol of relevant military police activities while wearing a single body-fixed tri-axial accelerometer and a heart rate monitor. For reference, all activities were recorded with a video camera and were manually labelled by one researcher. We created two sets of labels: one detailed set that incorporated all relevant activities (sprinting in a straight line, sprinting in a zigzag) and one lean set that used broader labels (i.e. sprinting). We extracted key components with a principal component analysis and used those to train a random forest classifier. We internally validated the model using the leave-one-subject-out method to create a confusion matrix for both label sets against the reference method, including their accuracies.

Results: The lean label set was predicted with 82.8% accuracy, which reflected a moderate to strong classifier. The detailed label set was predicted with 63.2% accuracy. This reflected a weak classifier due to misclassifications of short non-repetitive activities. The analysis further revealed that heart rate was a non-distinctive feature in activity recognition for all activities.

Discussion: The algorithm developed in this study accurately predicted continuous repetitive activities well, but not short non-repetitive activities. The results of this study go beyond previous reports, showing that the classification model of the lean label set could classify carrying, crawling and digging in a military setting with moderate accuracy. Future studies may incorporate multiple sensors and more features to better distinguish between short non-repetitive activities.

Conclusion: Using raw data from one tri-axial body-fixed accelerometer, the algorithm developed in this study could predict specific military police activities with 63.2% accuracy and general activities with 82.8% accuracy. More research is needed to classify short non-repetitive activities accurately.