Department of Advanced Convergence, BK21 FOUR, Handong Global University, Pohang 37554, Republic of Korea.
College of ICT Construction & Welfare Convergence, Kangnam University, 40, Yongin 16979, Republic of Korea.
Sensors (Basel). 2023 Jan 15;23(2):1003. doi: 10.3390/s23021003.
Resistance bands are widely used nowadays to enhance muscle strength due to their high portability, but the relationship between resistance band workouts and conventional dumbbell weight training is still unclear. Thus, this study suggests a convolutional neural network model that identifies the type of band workout and counts the number of repetitions and a regression model that deduces the band force that corresponds to the one-repetition maximum. Thirty subjects performed five different exercises using resistance bands and dumbbells. Joint movements during each exercise were collected using a camera and an inertial measurement unit. By using different types of input data, several models were created and compared. As a result, the accuracy of the convolutional neural network model using inertial measurement units and joint position is 98.83%. The mean absolute error of the repetition counting algorithm ranges from 0.88 (seated row) to 3.21 (overhead triceps extension). Lastly, the values of adjusted r-square for the 5 exercises are 0.8415 (chest press), 0.9202 (shoulder press), 0.8429 (seated row), 0.8778 (biceps curl), and 0.9232 (overhead triceps extension). In conclusion, the model using 10-channel inertial measurement unit data and joint position data has the best accuracy. However, the model needs to improve the inaccuracies resulting from non-linear movements and one-time performance.
如今，由于阻力带便于携带，因此被广泛用于增强肌肉力量，但阻力带锻炼与传统哑铃重量训练之间的关系仍不清楚。因此，本研究提出了一种卷积神经网络模型，用于识别带锻炼类型和计数重复次数，以及一种回归模型，用于推断与最大重复次数相对应的带力。30 名受试者使用阻力带和哑铃进行了 5 种不同的练习。使用相机和惯性测量单元收集了每种运动的关节运动。通过使用不同类型的输入数据，创建并比较了几个模型。结果，使用惯性测量单元和关节位置的卷积神经网络模型的准确率为 98.83%。重复计数算法的平均绝对误差范围为 0.88（坐姿划船）至 3.21（过头三头肌伸展）。最后，5 种练习的调整后 r 方值分别为 0.8415（卧推）、0.9202（肩推）、0.8429（坐姿划船）、0.8778（二头肌弯举）和 0.9232（过头三头肌伸展）。总之，使用 10 通道惯性测量单元数据和关节位置数据的模型具有最高的准确率。但是，该模型需要改进由于非线性运动和一次性表现导致的不准确性。
