Recognizing human actions from still images

In this study, a new method is proposed for recognizing actions from still images with general content. High increase in the number of action images in recent years bring about the need for indexing and large-scale visual search for these images. Although some previous researches mainly worked on body pose estimation, it is still challenging to estimate body pose from daily activity images like interacting with computer, photographing and walking, as viewpoint variability and human pose occlusion occurs a lot in them. Following the idea of the correlation between human action, objects likely to occur in the action image, human factor and scene type, Spatial Pyramid representation is used. Besides local matching performed by Scale Invariant Feature Transform, the performance of a global representation which extracts spectral features of the image, namely Gist descriptors are also investigated. Considering image representation, a binary classifier, Support Vector Machines is trained using 1-vs-all method for multiclass classification and test results are shown.