In this paper, we present a novel human action recognition (HAR) framework from depth motion maps and the constructed feature maps from the Kinect skeletal data namely joint angular displacement maps (JADM’s) inputted to a novel Granular convolutional neural network (GCNN). To represent a human action our GCNN requires two inputs. First input is a depth optical flow maps (DOFM) obtained by accumulating the optical flow frames from depth images consequently. Whilst, the second input is a JADM’s constructed from the Kinect skeletal data. To increase the recognition rates of HAR, we are proposing the GCNN architecture which is having two inputs one with DOFMs and other with JADMs were both the streams are concatenated separately to form a third stream. Now these streams are finally fused for the recognition tasks of the actions. To investigate the performance of our model we are testing with our own created data BVCAction3D and with publicly available datasets MSR daily activity 3D, UT Kinect and NTU RGB-D.
Volume 11 | 06-Special Issue
Pages: 1210-1219