Two dimensional markerless gait analysis protocol for estimating the sagittal lower limb joint kinematics with a single RGB-D camera for clinical applications.

The measure of lower limb joint kinematics is helpful in evaluating gait alterations. Nowadays, optoelectronic stereophotogrammetry gait analysis is the gold-standard in clinical practice for quantitative assessment of human motion because it provides accurate results, but it is not without limitations. Video-based markerless (ML) systems can represent a promising alternative to 3D marker-based systems because it allows to make the experimental sessions faster and simpler because does not require the application of fixtures on the skin of the patients. In addition, recent advancements in depth camera (RGB-D) technology have paved the way to the development of a new generation of low-cost movement analysis systems. The aim of the thesis project is the development of a two-dimensional markerless (2D ML) gait analysis protocol through the use of a single RGB-D camera (RGB camera with a depth infrared sensor) to estimate the lower limb joint kinematics for clinical applications. The proposed protocol for the estimation of the joint angles consists in four parts: -??Image segmentation: separation of the subject from the background; -??Multi-segmental model: creation of a lower limb model; -??Joint center tracking: the trajectories lower limb anatomical landmarks were estimated by using the ICP algorithm; -??Estimation of the joint kinematics: the orientations of the lower limb segments were determined from joint center trajectories. This thesis project is divided in four sections. The first one aims to implements an automatic segmentation algorithm to separate the subject from the background without the need of a homogeneous green background during the experimental session. The image segmentation is a key component in a lot of current markerless gait analysis systems. It represents a difficult step because the automated choice of an optimal threshold to separate the subject from the background is challenging if the background is inhomogeneous and if the background has similar colours to some body regions of the analysed subject. The presence of a coloured background is required to simplify this first step because it standardises the experimental scenario, but it complicates the experimental set-up conditions by increasing the number of required tools to obtain a 2D ML gait analysis. To achieve this goal, a modified version of Sal et al. was implemented to obtain the subject separation from the background. Subsequently, the algorithm relative performances are assessed by comparing the obtained segmentations with manual masks by calculating the Jaccard index. The second part of this work aims to apply a modified version of the algorithms proposed in Pantzar-Castilla et al. to extract the foreground lower limb and to estimate the lower limb joint kinematics. The third part is a feasibility study in order to improve the quality of the obtained results by evaluating how the light condition affects them and by studying the performances of the RGB-D cameras used to capture the video. To achieve this goal, an experimental session was conducted and subsequently the experimental approach is presented: it starts to acquire video with two different depth cameras in order to select the hardware that offers the higher image quality. Subsequently, the proposed algorithms were implemented. Finally, the obtained joint kinematics, in different setup conditions, are compared with the ones obtained starting from data of stereophotogrammetry system.