Advanced Machine Learning Techniques to Analyze Video Recordings for Application in Rehabilitation Medicine

The gait function has an important impact on the quality of life of people. Nowadays numerous pathologies are the cause of a deviation from the normal gait, resulting in a great disability in the performance of the daily life activities. For this reason, the improvement of gait is one of the main focuses of rehabilitation interventions to ensure total inclusion in society and to remove every kind of barrier. Currently, motion analysis laboratories perform gait analysis using systems that exploit the tracking of infrared reflective markers positioned on anatomical landmarks of the patient. Unfortunately, these systems are cumbersome, high-cost and take a lot of hours to collect and to analyze the clinical walking data by specialized personnel. The purpose of the thesis aims to evaluate a new marker-less approach to perform gait analysis on disabled patients by enabling low-cost, and user-friendly procedures. These techniques are based on video recordings and the use of advanced machine learning techniques to derive the position of the human landmarks used to extract the biomechanical data of the gait. In this work, a preliminary assessment on the reliability of Openpose, a motion tracking algorithm, is performed to define the limits and to evaluate whether it will be possible the use for gait analysis. An algorithm for the extraction of the main angles of the lower limb is developed in the Matlab environment and some Openpose limits are addressed. The comparison between Openpose, the modern tracking algorithm, and the Vicon system, a traditional motion analysis system, is carried out by studying the gait parameters of five healthy subjects. The main objectives of this work are the comparison of the two biomechanical models, in terms of body segments-orientation and joint-angles, the comparison of the position of the centers of rotation estimated by the systems and the characterization of the Openpose estimated error in terms of angles shape, angles magnitude and markers displacement. The final results obtained show a good estimate of the gait parameters with the use of cameras only, reporting errors in estimating the knee and the ankle angles of less than 5 degrees, and errors in estimating the hip angle up to 10 degrees. The preliminary evaluation of Openpose highlights the good potential for future use in the clinic of the pose estimation tool. Nevertheless, certain limits must be further exceeded to the introduction in the evaluation of patients with severe levels of disability.