The purpose of research is to analyze the gait parameters of figure in the video when he is not 90 degrees to the camera position through the trained SemGCN model, and analyze its reliability. Two groups of video data in the experiment are collected in the same time, meaning experimental group and reference group. In the video data of experimental group, the angle between figure and camera position is 120 degrees, while it is 90 degrees in the video data of reference group. The 2D joint point coordinate information of video data in experimental group is obtained firstly through 2D network, and then it is delivered to trained SemGCN model to regress 3D model and extract 3D joint point coordinate information, then output knee angle curve, which would be processed for filtering and smoothing by Savitzky-Golay wave filter. The video data in reference group is directly processed through Openpose algorithm, and the knee angle curve is output. Finally, the relevance of data in two groups are analyzed by adopting intraclass correlation coefficient(ICC). The result shows that the correlation of a single measurement is 0.898, meaning the data in two groups are strongly correlated. The reliability of gait parameters of figure in the video which are analyzed by using SemGCN model is good, which can be applied for the measurement of figure joint angle in video or other relevant fields.