Human Motion Capture Using a Multi-2D Pose Estimation Model
Guitao Cao, Yihao Pu, Yan Li, Zhenwei Zhao · 2019
In this paper, we raise a method to find certain motion segments in the video using a Multi-2d pose estimation model. Several types of angles are extracted. We draw a continuous curve based on the entire video for all required angle values and smooth the angle curve to reduce the influence of human detection errors on motion recognition. After the smoothing process, the curve is again processed into a change curve as analysis and then search for two kinds of movements with a reasonable threshold. We also handle the noise in the recognition result to get more accurate motion fragments. By comparing and analyzing different parameters, the optimal parameters to achieve lower error is found. We coded the system and carried out a large number of experimental analysis mainly through the analysis of the characters in the video to turn around and get up, and the following results are achieved: 1)Be able to use effectively optimized data for effective analysis. 2)In a multi-person situation, the classification of each person's data based on time series is achieved. 3)Fragments of two motion states are detected more accurately.