A comparative evaluation of two markerless methods for sagittal lower limb joint kinematics estimation based on a single RGB-D camera

Gait analysis is a valuable and widespread tool for assessing the quality of human locomotion. To date, marker-based optoelectronic stereo-photogrammetry (MB) represents the gold standard for the evaluation of lower limb joint kinematics and the most accurate solution. However, MB systems require specialized operators, dedicated spaces, and long preparation and processing time. Additionally, the presence of markers attached to the subject’s body and a cumbersome set-up of many cameras may affect the spontaneity of the movement. Video-based markerless systems (ML) represent a low-cost, powerful, and promising alternative to MB systems. The use of ML techniques allows making the experimental sessions faster and easier since it does not require the application of markers on the skin of the patients. Most recently, several companies are producing inexpensive tracking consume electronics systems constituted by an RGB camera integrated with an infrared depth sensor (RGB-D camera). These cameras (e.g. Microsoft Kinect, IntelRealSense D435) often come with software development kits (SDK) for real-time tracking of body position and orientation primarily focused on gaming purposes. Very recently (2020), a new RGB-D camera (Azure Kinect) was released by Microsoft and compared to the previous versions of Kinect, this camera is targeted towards other markets such as logistics, robotics, health care, and retail. The improved performances suggest the possibility to apply these technologies for the development of clinical-based applications. Within this general context, this thesis project aims: (i) to investigate whether motion tracking through the body tracking SDK integrated into the Azure Kinect DK could be employed to perform gait analysis for clinical purposes and (ii) to compare the performances of the above-mentioned SDK to an improved custom version of a 2D markerless method (MLM) based on a subject-specific kinematic model developed by Balta et al., 2020. To assess the performance of the methods, the sagittal lower limb joint kinematics were validated against a standard marker-based (MB) gait analysis protocol. Five healthy subjects were acquired in a gait analysis laboratory equipped with the Azure Kinect placed laterally to the walkway and an MB system (Vicon Vero) that was used as the gold standard. The acquisitions could not be synchronized due to the mutual interference of the two systems caused by the same working IR wavelength. Seven significant gait variables were extracted for each trial: the knee flexion at initial contact, its max value during the loading response and the swing phase and its max extension in the stance phase; the ankle max dorsiflexion during the stance and swing phase and the hip max extension during the stance phase. Results were compared by calculating the mean differences obtained from the SDK and the markerless method (MLM) with respect to the MB and their reliability was evaluated with the intraclass correlation coefficient (ICC). Both the SDK and MLM provided good estimates for knee and hip kinematics while higher differences were found for ankle kinematics, mainly for the SDK in which the foot segment was identified from the lateral malleolus to the toe position. Although the MLM is more time-consuming than SDK, its accuracy overcomes the SDK limitations, and it could be further improved by implementing a 3D model which could embed the outputs of both the SDK and the MLM method.