2DHOOF-2DPCA contour based optical flow algorithm for human activity recognition

A novel algorithm for human activity recognition is presented in this paper. This approach is based on a new 2D representation for the Histogram of Oriented Optical Flow (2DHOOF) describing the motion of the actor´s contour, where one multi-layer 2D-histogram per video is constructed. Each histogram layer consists of 2D bins (layers) that represent different range of angles. Applying our 2DHOOF features descriptors on the actor´s contour reduces the storage requirement and the computation complexity since a sparse optical flow is calculated instead of dense optical flow. In addition, it is robust to variations in the background, actor´s appearance, and imperfections in actor´s contour. This new 2D representation allows the usage of the Two Dimensional Principle Component Analyses (2DPCA) which maintains the spatial relation of the motion, and provides further high accuracy and low computation complexity. Experimental results applied on the Weizmann and IXMAS datasets achieved the highest reported recognition accuracy and the fastest runtime compared to recent methods.