Early Late Fusion of features for Skeleton based Human Action Recognition using Adaptive Kernel Learning

The identification of human action or behavior is very much discussed through the use of computer vision and artificial intelligence. Various action recognition strategies were proposed by many researchers to extract the action features in order to identify the action accurately. In the present article, we developed a new geometric feature fusion framework like score and features fusion on 3D skeletal data. First, geometric features like relative distance features, relative angle features are extracted from skeleton data. Next, various fusion models were adopted in early fusion combining features and late fusion after classifier to combine the output scores of both geometric features. In this work, the extracted features trained with the adaptive kernel learning classifier for recognizing actions. To test the working of our architecture, tested with other architectures on our own PACE3DAction and benchmark skeleton action datasets like PKU-MMD and NTU RGB-D.