Using deep learning-based pose estimation algorithms for markerless gait analysis in rehabilitation medicine

Walking is one of the most natural human activities and certainly the most impactful on one's quality of life. However, the human ability to walk can be compromised by neurological, orthopedic, or traumatological factors. When the gait is impaired by one or multiple of these factors, a key objective of modern rehabilitation interventions is to help people with gait impairments to regain or improve their ability to walk, minimizing the negative impact on their quality of life at both a social and personal level. Nowadays, gait analysis laboratories use multiple technologies to evaluate and improve the individual's gait patterns. Among the various tools available for gait analysis, Motion Capture systems based on cameras and infrared-reflective markers positioned on the individual are used as gold standard, due to their high accuracy. However, these systems have the disadvantage of being cumbersome, costly, complex, and not time-efficient. The proposed study aims at using a new motion detection approach that relies on deep learning-based algorithms with the long-term goal of providing a simple, cost and time-effective alternative to motion analysis systems currently used in rehabilitation medicine. In this study, five healthy subjects without any motor disabilities were recruited and asked to walk on a treadmill at different speeds to collect video data. The collection of this data has made it possible to carry out the biomechanical analysis of movement and the estimation of biomechanical parameters of clinical interest both using the gold standard gait analysis system and using the modern system based on machine learning. All this made it possible to compare the gait parameters extracted through the two tracking motion systems, OpenPose, and Vicon. The objective of the thesis is the comparison of these two motion tracking tools with particular focus on the joint angles of the lower limbs, specifically the angles of the hip, knee, and ankle, on the position of the centers of joint rotation and on the fundamental kinematic parameters of the gait cycle. We quantify the error in terms of the shape of the trajectories of the body joints, their displacement, and the shape and magnitude of the angles. To cope with the purpose of this work, there are preliminary stages of processing the data extracted from the two systems. A simplified biomechanical model was developed to allow the calculation of the angles characterizing the lower limb through the data extracted from the OpenPose system. The obtained results show a high correlation between the hip and knee angles extracted with the two systems with a Root Mean Square Error (RMSE) below 5 degrees and a Mean Absolute Error (MAE) below 4 degrees. The comparison between the ankle angles obtained by the two systems show an RMSE below 8.5 degrees, an MAE below 6.5 degrees, and a correlation index greater than 0.5 for all the subjects analyzed. The cause for the major error obtained in the ankle angle is due to an inaccurate estimate of the foot key-points during the subject's walking by the OpenPose pose estimation tool. In conclusion, it can be said that the OpenPose system has great potential for its application in rehabilitation medicine. However, further investigation and improvements are needed to make it more robust and to allow its use even with people who have motor disabilities.