Key Posture Extraction Method using a Small Number of Motion Datasets
Soungsill Park, Young Ho Chai · Moving Image & Technology (MINT) · 2023
Although human movement can be inferred from various datasets, a method is required to recognize the change in movement from a small amount of new data. A deep learning-based method cannot be used to learn motion from limited data. To use existing feature extraction methods, a process is required to extract key postures from the entire dataset. The proposed method extracts key postures from an entire frame, whereby the differences in distance from the starting posture are added and displayed as a graph. The key task is to determine an inflection point in the graph, divide the data, and designate the frame corresponding to the inflection as an intermediate frame. Kinect and inertial sensor data are used to demonstrate the results of the proposed method. The proposed method can be used to determine similar motions on the basis of key postures, using existing datasets.