Quantifying swimming activities using accelerometer signal processing and machine learning: A pilot study

Recognition of TaeKwonDo kicking techniques based on accelerometer sensors

TaeKwonDo (TKD) is a worldwide sport in both competitive among athletes and physical exercise among the public scenarios. Measuring TKD kicks have been studied a lot in a laboratory setting but rarely in a free-living situation. Machine learning algorithm combined with accelerometer data was used to study some martial art styles, e.g., Chinese KungFu but little in TKD. The purpose of this study was to discover a method to recognize different kicking techniques in TKD. A total of 20 participants (35 % male) were recruited to perform front kick, roundhouse kick, side kick and back kick 6 times on each side with three accelerometers wore on waist, right ankle and left ankle. SVM and decision tree were used to analyze data and classify kicking movements. The usage of different combination of accelerometers were also compared. The result showed that using accelerometers on waist and both ankles, on waist and only right ankle, on only waist and combined with SVM model could have at least 0.96 accuracy of classification, while decision tree had the accuracies around 0.8. It was concluded that using SVM model on only waist data is the optimal choice because of the high accuracy and less accelerometers used.

Accelerometer-based prediction of ground reaction force in head-out water exercise with different exercise intensity countermovement jump

Background

Water jumping exercise is an alternative method to achieve maintenance of bone health and reduce exercise injuries. Clarifying the ground reaction force (GRF) of moderate and high cardiopulmonary exercise intensities for jumping movements can help quantify the impact force during different exercise intensities. Accelerometers have been explored for measuring skeletal mechanical loading by estimating the GRFs. Predictive regression equations for GRF using ACC on land have already been developed and performed outside laboratory settings, whereas a predictive regression equation for GRF in water exercises is not yet established. The purpose of this study was to determine the best accelerometer wear-position for three exercise intensities and develop and validate the ground reaction force (GRF) prediction equation.

Methods

Twelve healthy women (23.6 ± 1.83 years, 158.2 ± 5.33 cm, 53.1 ± 7.50 kg) were recruited as participants. Triaxial accelerometers were affixed 3 cm above the medial malleolus of the tibia, fifth lumbar vertebra, and seventh cervical vertebra (C7). The countermovement jump (CMJ) cadence started at 80 beats/min and increased by 5 beats per 20 s to reach 50%, 65%, and 80% heart rate reserves, and then participants jumped five more times. One-way repeated analysis of variance was used to determine acceleration differences among wear-positions and exercise intensities. Pearson’s correlation was used to determine the correlation between the acceleration and GRF per body weight on land (GRF_VL_BW). Backward regression analysis was used to generate GRF_VL_BW prediction equations from full models with C7 acceleration (C7 ACC), age, percentage of water deep divided by body height (PWDH), and bodyweight as predictors. Paired t-test was used to determine GRF_VL_BW differences between values from the prediction equation and force plate measurement during validation. Lin’s CCC and Bland–Altman plots were used to determine the agreement between the predicted and force plate-measured GRF_VL_BW.

Results

The raw full profile data for the resultant acceleration showed that the acceleration curve of C7 was similar to that of GRFv. The predicted formula was − 1.712 + 0.658 * C7ACC + 0.016 * PWDH + 0.008 * age + 0.003*weight. Lin’s CCC score was 0.7453, with bias of 0.369%.

Conclusion

The resultant acceleration measured at C7 was identified as the valid estimated GRF_VL_BW during CMJ in water.