articles

The problem of recognizing human body activities by analyzing data obtained from inertial measurement units using artificial neural networks is consired. The feature of this study is using of only five angular coordinates defined by a planar five-link human model for movement analysis. Furthermore, in addition to walking, athletic exercises movements are analyzed. Both long-term and short-term movements (for example, Start of walking or sitting down) are recognized. Activities recognition is used in human-machine interfaces, as well as in sports, to assess the number and technique of performing exercises. Movement analysis is also used in medicine, for example, in the treatment of Parkinson’s disease. The purpose of this work is to develop a method for recognizing short-term and long-term movements based on data from wearable inertial sensors, without the influence of vertical axis body position, with a recognition accuracy of over 98%. Five measuring modules are attached to the human body using elastic belts. One module is placed on the person’s chest, the others are attached to the sides of the thighs and shins. Each module uses a microelectromechanical sensor MPU-6050 with a specialized digital signal processor for motion processing (DMP) used to pre-process and filter data coming directly from the 3-axis gyroscope and accelerometer. Then the data from the modules is transmitted wirelessly to the receiving device and then to a personal computer, where they are finally processed and stored for subsequent classification. In the course of the work, a data set was collected that, in addition to common states (standing and sitting positions, walking), included sports exercises that primarily involve the leg muscles (squatting, step-ups). Using video recording carried out during the motion capture process, the areas where the movements are performed were marked on the obtained data. Several neural network configurations (fully connected, convolutional, and recurrent) were used to classify the movements. The highest accuracy of combine recognition of long-term and short-term movements (98–99% F1-score) was achieved using convolutional neural networks. In the work process, it was found that the duration of the analyzed movement fragment has effect on recognition accuracy. This accuracy was achieved using 2-second recording fragments.