Contributions to physical exercises monitoring with inertial measurement units

Abstract

Resumen: La monitorización de movimientos trata de obtener información sobre su ejecución, siendo esencial en múltiples aplicaciones, como el seguimiento de terapias físicas. La monitorización tiene un doble objetivo esencial para lograr los beneficios de dichas terapias: asegurar la corrección en la ejecución de movimientos y mejorar la adherencia a los programas prescritos. Para lograr esta monitorización de forma remota y poco intrusiva, se necesitan recursos tecnológicos. Este trabajo se centra en las soluciones basadas en sensores inerciales. Esta tesis estudia los algoritmos de la literatura para el análisis de movimientos con sensores inerciales, determinando un parámetro anatómico requerido en diversas propuestas: la posición de las articulaciones respecto de los sensores, así como longitud de los segmentos anatómicos. En este trabajo se introducen dos algoritmos de calibración anatómica. El primero, basado en mínimos cuadrados, determina el punto o el eje medios de aceleración nula presente en las articulaciones fijas. El algoritmo está adaptado a los movimientos lentos dados en los miembros inferiores para estabilizar las articulaciones. El segundo, adaptado a la variación de la posición relativa del punto de aceleración nula respecto de los sensores a causa del característico tejido blando asociado al cuerpo humano, emplea las medidas inerciales como entradas en un filtro de Kalman extendido. Por otro lado, esta tesis aborda la falta de datos comunes para la evaluación y comparación de los algoritmos. Para ello, se diseña y crea una base de datos centrada en movimientos habituales en rutinas físicas, que se encuentra publicada en Zenodo. Esta base de datos contiene movimientos de calibración articular y de ejercicios de miembros inferiores y superiores ejecutados de forma correcta e incorrecta por 30 voluntarios de ambos sexos con un amplio rango de edades, grabados con cuatro sensores inerciales y un sistema de referencia óptico de alta precisión. Además, las grabaciones se encuentran etiquetadas acorde al tipo de ejercicio realizado y su evaluación. Finalmente, se estudia un segundo enfoque de monitorización de rutinas físicas, cuyo objetivo es reconocer y evaluar simultáneamente los ejercicios ejecutados, retos comúnmente estudiados individualmente. Se proponen tres sistemas que emplean las medidas de cuatro sensores inerciales y difieren en el nivel de detalle en las salidas del sistema. Para realizar las clasificaciones propuestas, se evalúan seis algoritmos de machine learning determinando el más adecuado.

This thesis is framed in the field of remote motion monitoring, which aims to obtain information about the execution of movements. This information is essential in many applications, including the clinical ones, to measure the evolution of patients, to assess possible pathologies, such as motor or cognitive ones, and to follow up physical therapies. The monitoring of physical therapies has twofold purpose: to ensure the correct execution of movements and to improve adherence to the programs. Both purposes are essential to achieve the benefits associated with physical therapies. To accomplish this monitoring in a remote and non-intrusive way, technological resources such as the well-known inertial sensors are needed, which are commonly integrated into the so-called wearables. This work focuses on inertial-based solutions for monitoring physical therapy routines. However, the results of this work are not exclusive of this field, being able to be applied in other fields that require a motion monitoring. This work is intended to meet the needs of the monitoring systems found in the literature. In the review of previous proposals for remote monitoring of rehabilitation routines, we found two different main approaches. The first one is based on the analysis of movements, which estimates kinematic parameters, and the second one focuses on the qualitative characterization of the movements. From this differentiation, we identify and contribute to the limitations of each approach. With regard to the motion analysis for the estimation of kinematic parameters, we found an anatomical parameter required in various methods proposed in the literature. This parameter consists in the position of the joints with respect to the sensors, and sometimes these methods also require the length of the anatomical segments. The determination of these internal parameters is complex and is usually performed in controlled environments with optical systems or through palpation of anatomical landmarks by trained personnel. There is a lack of algorithms that determine these anatomical parameters using inertial sensors. This work introduces an algorithm for this anatomical calibration, which is based on the determination of the point of zero acceleration present in fixed joints. We use one inertial sensor per joint in order to simplify the complexity of algorithms versus using xv xvi ABSTRACT more than one. Since the relative position of this point may vary due to soft tissue movements or joint motion, the mean null acceleration point for the calibration motion is estimated by least squares. This algorithm is adapted to slow movements occurring in the lower-limbs to meet the required joint stabilization. Moreover, it can be applied to both joint centers and axes, although the latter is more complex to determine. Since we are dealing with the calibration of a system as complex as the human body, we evaluate different movements and their relation to the accuracy of the proposed system. This thesis also proposes a second, more versatile calibration method, which is adapted to the characteristic soft tissue associated with the human body. This method is based on the measurements of one inertial sensors used as inputs of an extended Kalman filter. We test the proposal both in synthetic data and in the real scenario of hip center of rotation determination. In simulations it provides an accuracy of 3% and in the real scenario, where the reference is obtained with a high precision optical system, the accuracy is 10 %. In this way, we propose a novel algorithm that localizes the joints adaptively to the motion of the tissues. In addition, this work addresses another limitation of motion analysis which is the lack of common datasets for the evaluation of algorithms and for the development of new proposals of motion monitoring methods. For this purpose, we design and create a public database focused on common movements in rehabilitation routines. Its design takes into account the joint calibration that is usually considered for the monitoring of joint parameters, performing functional movements for it. We monitor lower and upper limb exercises correctly and incorrectly performed by 30 volunteers of both sexes and a wide range of ages. One of the main objectives to be fulfilled by this database is the validation of algorithms based on inertial systems. Thus, it is recorded by using four inertial systems placed on different body limbs and including a highly accurate reference system based on infrared cameras. In addition, the recorded movements are labeled according to their characterization, which is based on the type of exercise performed and their quality. We provide a total of 7 076 files of inertial kinematic data with a high-precision reference, characterized with respect to the kind of performed motion and their correctness in performance, together with a function for automatic processing. Finally, we focus on the analysis of the second approach of monitoring physical routines, whose objective is to obtain qualitative information of their execution. This work contributes to the characterization of movements including their recognition and evaluation, which are usually studied separately. We propose three classification systems which use four inertial sensors. The proposals differ in the distribution of data and, therefore, the level of detail in the system outputs. We evaluate six machine learning techniques for the proposed classification systems in order to determine the most suitable for each of them: Support Vector Machines, Decision Trees, Random Forest, xvii K Nearest Neighbors, Extreme Learning Machines and Multi-Layer Perceptron. The proposals result in accuracy, F1-value, precision and sensitivity above the 88 %. Furthermore, we achieve a system with an accuracy of 95% in the complete qualitative characterization of the motions, which recognizes the performed motion and evaluates the correctness of its performance. It is worth highlighting that the highest metrics are always obtained with Support Vector Machines, among all the methods evaluated. The proposed classifier that provides the highest metrics is the one divided into two stages, that first recognizes the exercises and then evaluates them, compared with the other proposals that perform both tasks in one single-stage classification. From our work, it can be concluded that inertial systems are appropriate for remote physical exercise monitoring. On the one hand, they are suitable for the calibration of human joints necessary for various methods of motion analysis using one inertial sensor per joint. These sensors allow to obtain the estimation of an average joint location as well as the average length of anatomical segments. Also, joint centers can be located in scenarios where joint-related sensor movements occur, associated with soft tissue movement. On the other hand, a complete characterization of the physical exercises performed can be achieved with four inertial sensors and the appropriate algorithms. In this way, anatomical information can be obtained, as well as quantitative and qualitative information on the execution of physical therapies through the use of inertial sensors.