A Salient Invariant Feature Descriptor For Human Action Recognition

ABSTRACT

The dramatic progress of studies in human action recognition has being attributed to challenges inherent with conventional methods such as bag-of-words based description. As a result, researchers in the field of computer vision are still making efforts towards achieving structured interpretation of complex activities between multiple objects. This study proposes Pyramid of Histogram Oriented Gradients (PHOG) computed from Depth Motion Maps in a video stream as a new feature descriptor for recognition of human activities. The proposed method has two steps which includes construction of depth motion maps from frames in sequence of a given video, and representation of images in each frame using PHOG. The latter step reflects local shapes and spatial layouts of images in three views of the depth maps. 1_2- regularized Collaborative Representation Classifier was adopted for classification of human activities in the processed depth images. Performance of the proposed invariant method was evaluated by using the MSR Action3D dataset, and compared ~ith that of conventional methods. Our result shows that our novel invariant feature descriptor improves the average rate of activity recognition with up to 12.52%with respect to conventional methods.