FocusRec RNN: Enhancing 3D Human Pose Reconstruction and Prediction with Wearable Motion Capture Technology
S. V. Evangelin Sonia, C. Beulah Christalin Latha, R. Venkatesan, Gnanadesigan Naveen Sundar, C. P. Shirley, Eleanor Stuart · 2024
Wearable motion capture technology has advanced significantly, enabling detailed and accurate tracking of human movements in three-dimensional space. This paper introduces FocusRec RNN, an innovative Recurrent Neural Network (RNN) architecture designed to enhance the reconstruction and prediction of 3D human poses using wearable IMU sensors. By incorporating a focus mechanism that dynamically prioritizes critical time-steps and body segments, FocusRec RNN improves upon traditional methods in terms of accuracy and robustness. Evaluation on datasets such as Human3.6M and the CMU Motion Capture Database demonstrates that FocusRec RNN achieves lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), alongside higher Percentage of Correct Keypoints (PCK) compared to traditional RNNs, LSTM networks, and CNNbased approaches. These improvements highlight FocusRec RNN’s potential for applications in sports performance analysis, medical diagnostics, rehabilitation, and interactive entertainment.