MotionNeXt: A Deep Learning Neural Network for Recognizing Human Motions Related Activities Using Inertial Sensors

Recognition of human activities using wearable sensors is essential for applications related to health and well-being. However, the complex data captured by inertial sensors poses challenges in accurately identifying actions. This research introduces MotionNeXt, a residual deep neural network incorporating aggregated transformation. MotionNeXt utilizes residual and multi-kernel blocks to extract spatial and temporal characteristics from raw IMU data. Subsequently, it employs global average pooling, a fully connected layer, and softmax for classification purposes. An attention layer enables the model to focus on specific segments of the input sequence while categorizing each time step. MotionNeXt is evaluated using a publicly available dataset of activities collected from wearable inertial measurement devices. It achieves state-of-the-art accuracy, boasting an F1-score of 99.37%, surpassing previous deep learning methods by 2- 5%. By integrating accelerometer, gyroscope, and magnetometer modalities, MotionNeXt mitigates the limitations of individual inputs, significantly improving accuracy across various models.