Physical Activity Data Acquisition and Processing via Inertial Sensor Networks

MSc Thesis

This thesis examines the use of wearable inertial sensors (Inertial Measurement Units - IMUs) for recording and analysing motion, with the aim of supporting rehabilitation programs and preventing pathological conditions such as deep vein thrombosis (DVT). The objective of the thesis is the development of a comprehensive infrastructure that combines accurate motion data collection with visualization and machine learning techniques, in order to provide a reliable assessment of human movement.

For this purpose, a comparative study and analysis of commercially available personal motion sensors (Inertial Measurement Units - IMUs) was conducted, based on their technical characteristics, ergonomics, cost, availability, and compatibility with medical standards. From this evaluation, the Movella DOT® sensor emerged as a suitable choice, as it successfully combines accuracy, portability, and software support.

Subsequently, a software infrastructure was developed in Python using the supplier’s Software Development Kit (SDK), based on the following stages: (a) collection and synchronization of lower-limb motion data in real time, (b) analysis and processing of the data, (c) three-dimensional visualization of lower limb motion, and the accession of a machine learning algorithm for the automatic recognition and correction of errors during the execution of motion exercises.

Data were collected by placing three sensors (on the thigh, the shin, and the dorsum of the lower limb) on a healthy subject during the execution of five predefined exercises, selected with an emphasis on the prevention of DVT. The analysis was based on Euler angles and was used to calculate the positions of the hip, knee, and ankle joints. In addition, the Random Forest algorithm was applied, trained using the aforementioned data, achieving satisfactory accuracy in detecting deviations from normal motion patterns.

The findings demonstrated that the proposed infrastructure is capable of accurately reproducing lower limb motion through three-dimensional visualization, while simultaneously identifying and highlighting errors in exercise execution. This approach enhances the reliability of assessment under real time conditions and enables immediate feedback.

In conclusion, a fundamental laboratory infrastructure was created for motion capture using wearable sensors, with the aim of signal analysis through machine learning models. Thus, a foundation is laid for future research and improvements, which may include the use of more extensive and heterogeneous datasets, as well as the incorporation of additional sensors and machine learning models, targeting a multidimensional and comprehensive analysis of human motion.