Action Recognition
Action Recognition
Action Recognition Based on Multi-Level Representation of a Space Time Shape
This approach considers human actions as space time shapes which is a concatenation of silhouettes, and uses a combination of shape descriptors to extract features suitable for recognition. The current algorithm uses a 3D Euclidean distance transform as one shape descriptor, the purpose of which is to give interior values in accordance to the shape boundary which is different for different human actions. By using the values obtained, the space time shape is dissected inwards to get coarser and coarser representation. At each level of the coarser representation, the R-Transform and R-Translation which is a variant of the Radon transform, is extracted and these provide the action features. R-Transform gives the variation of the posture of the human silhouette at every instant of the space time shape while the R-Translation discriminates between actions which involves a large translation of the body with those which does not. The figure below gives the flow diagram of the algorithm along with sample space time shapes of jumping jack and walking.
Jumping Jack ST Walking ST
Distance Transform Values
Gradient of the Distance Transform
Segmentation of the Space Time Shape
R-Transform Feature
Video Demonstration
Action Recognition
(Research Demonstrated at Old Dominion University Vision Lab)
Segmentation for Human Activity Recognition