A Kalman Filter based Approach for Markerless Pose Tracking and Assessment

Monocular 3D Human Pose Markerless Systems for Gait Assessment

“…However, the innovation trends in computer vision and machine learning techniques are expected to provide better models for human pose estimation, which can cope with depth estimation in 2D video. The experiment results also suggest the previous gait processing solution [15] has defects that caused the processed gait signal to suffer the worst PE and DTW distance. However, KF with two optimisations can significantly improve the 10.7% similarity of the predicted signals to the ground truth, without generating additional global offset compared with BlazePose's raw prediction.…”

“…The FDF uses Discrete Fourier Transform (DFT) and Inverse Discrete Fourier Transform, supported by a suitable filter strategy, to select principal frequency components to recover the filtered signal. KF and FDF were both used in earlier work [ 15 ] to predict missing data and denoise the original gait signals. Studies have demonstrated that KF and FDF can help reduce assessment system failures caused by the above problems.…”

“…After the KF estimates the missing gait signals, the FDF using the Least Root Mean Square Error strategy (LRMSE) [ 15 ] will be used to calculate the principal frequency components and then recover them to the time domain gait signal. Additionally, previous research [ 15 ] suggested using the pose segmentation mask to reduce the background noise level. The KF and FDF algorithm complement and denoise original gait signals and return a processed gait signal.…”

“…In this paper, we propose a gait signal decoding function, which processes the data to produce discrete gait metrics, including joint flexion and striding speed. These discrete gait parameters are separately stored in different datasets in terms of the targets’ age and gender, which can later be used to create models for analysis, called “feature models” [ 15 ]. This function enables the assessment system to evaluate patients’ disease condition with a score for aiding healthcare workers, even non-specialists, to learn about patients’ recovery progress and make suitable clinical decisions.…”

“…In our previous work [ 15 ], we developed a markerless, non-invasive rehabilitation assessment system using a consumer monocular camera together with a human pose detection model, BlazePose [ 16 ], to assess the patient’s gait in an indoor environment that only requires a limited number of gait cycles (i.e., 2–3 strides for a video sample). Although there are other models such as OpenPose [ 17 ], D3KE [ 18 ] and 3DPoseNet [ 14 ] providing better accuracy in human pose estimation tasks, Refs.…”