Design of a non-invasive cost-effective mobile robotic system for human gait analysis optimized by machine learning algorithms

• Autores: Diego Andres Guffanti Martinez
• Directores de la Tesis: Alberto Brunete González (dir. tes.), Miguel Hernando Gutiérrez (codir. tes.)
• Lectura: En la Universidad Politécnica de Madrid ( España ) en 2021
• Idioma: español
• Materias:
  • Ciencias de la vida
    • Biofísica
      • Biomecánica
  • Ciencias tecnológicas
    • Tecnología de la instrumentación
      • Ingeniería de control
• Enlaces
  • Tesis en acceso abierto en: Archivo Digital UPM
• Resumen

  • This dissertation started from an original and clinically interesting question, namely, how can technology contribute to the objective assessment of the gait pattern of patients in clinical practice. In the search for an answer to this question, this dissertation tries to analyze the viability of using a cost-effective and non-invasive robotic-based system with depth cameras for human gait analysis, and to propose a new prototype. First of all, the current knowledge regarding the application of technology for the objectification of gait patterns has been mapped. On the basis of clear criteria, the possibilities of one or more depth cameras were investigated in different measuring set-ups. Three methods for using depth cameras in human gait analysis were analyzed: the first method proposed the use of one depth camera to analyze walking over a treadmill; the second one proposed the use of several depth cameras in a multi-sensor configuration, with the cameras placed in a row; and the third one proposed the integration of one depth camera on a mobile robot. The design of the robotic system was performed in two stages. In the first stage a straight-line follower robot for human gait analysis was designed and validated. The idea of this robot was to follow the human from the front while walking in a straight line. During the design stage several software related questions regarding model identification and configuration of the control law were addressed. The design of the control law required the integration of a lead compensator and a Filtered Smith Predictor (FSP) to compensate for sensor latency. Following the experience acquired in the preliminary studies, certain shortcomings of the depth sensors in the analysis of human gait were detected. Therefore, it became necessary to work on a sensor accuracy improvement stage. To improve the accuracy of the sensor neural networks were trained through supervised machine learning algorithms. The resulting accuracy of the system to retrieve kinematic gait data was validated by comparing results with a golden standard. The next step was to make the robot suitable for use in an actual practice setting, i.e. outside a lab environment. This required, besides mapping the environment and defining a trajectory, adequate tracking of both the trajectory and the subject. While addressing these requirements physical possibilities and constraints of the robot and the environment in which the robot should be able to function had to be taken into account. For this, newest insights in the field of environment mapping, lane keeping, and person following were applied. First clinical tests with patients with Multiple Sclerosis gave an initial impression of the applicability of the instrument in patients with abnormal walking patterns. Further validation followed in which different typical gait abnormalities were analyzed and results were compared to an inertial sensor system. During this latter study, the main differences between normal and pathological gait were identified based on joint kinematics and the main descriptors of gait. Parallel to the construction of the robot, an interface for robot operation, gait data management and post-processing was developed. This allowed medical staff to operate the mobile robotic system accurately, without the need for training sessions or technical expertise. To summarize, the main contributions of this thesis are: the scientific community has been provided with the main limitations and advantages of each configuration of gait analysis with depth sensors; the design, construction and validation of a mobile robotic system for human gait analysis; control architecture designed for person following; improvement of the accuracy of the depth sensor for human gait analysis through machine learning; the design of an application to operate the robot; and the identification of the main differences in normal and pathological patterns of gait.