Human action recognition from depth sensor via skeletal joint and shape trajectories with a time-series graph matching
D. Anil Kumar, E. Kiran Kumar, Mopidevi Suneetha, L. Rajasekhar, L. Rajasekhar · AIP conference proceedings · 2024
Action recognition (AR) has achieved outstanding performance in computer-vision techniques and machine learning in recent years.However, most of the other algorithms suffer from identifying the extract positions and shape of the action in video frame sequence.To work out the aforementioned problems, in this research paper implemented a new framework to extract specific features of trajectories and shape features within the joints in a 3D space, To utilize these positions and extracted shape features actions are classified by graph matching algorithm.The graphs are build using joint trajectories as the vertex and the measured relative joint distance metric as the edge for every frame in the action sequence.After that, the time series graph matching model is to find similarity between the action query within the actions.To analyze the implemented model to various state-of-the-art frameworks and NTU RGB-D, MSR Daily Activity, UT Kinect, and G3D are datasets have been used to test the proposed TSGM algorithm.