In this paper, we present techniques for automated understanding of tutor-student behavior through detecting visual deictic gestures, in the context of one-to-one mathematics tutoring. To the best knowledge of the authors, this is the first work in the area of intelligent tutoring systems, which focuses on spatial localization of deictic gestural activity, i.e. where the deictic gesture is pointing on the workspace. A new dataset called SDMATH is first introduced. The motivation for detecting deictic gestures and their spatial properties is established, followed by techniques for automatic localization of deictic gestures in a workspace. The techniques employ computer vision and machine learning steps such as GBVS saliency, binary morphology and HOG-SVM classification. It is shown that the method localizes the deictic tip with an accuracy of over 85 % accuracy for a cut off distance of 12 pixels. Furthermore, a detailed discussion using examples from the proposed dataset is presented on high-level inferences about the student-tutor interactions that can be derived from the integration of spatial and temporal localization of the deictic gestural activity using the proposed techniques.