Contents 1. Introduction
The goal of action recognition is to localize a particular event of interest in video, such as a tennis serve, both in space and in time. Just as object recognition is a key problem in image understanding, action recognition is a fundamental challenge for interpreting video. A recent trend in action recognition has been the emergence of techniques based on the volumetric analysis of video, where a sequence of images is treated as a three-dimensional space-time volume. Eschewing the building of explicit models of the actor or environment (e.g., kinematic models of humans), these approaches attempt to perform recognition directly on the raw video. An obvious benefit is that recognition need not be limited to a specific set of actors or actions but can, in principle, extend to a variety of events - given appropriate training data. The drawback is that volumetric representations do not easily generalize across appearance changes due to different actors, varying environmental conditions and camera viewpoint. This observation has motivated the employment of video features that are robust to appearance; these can be broadly categorized as shape-based (e.g., background subtracted human silhouettes) and flow-based (e.g., motion fields generated using optical flow). However, as discussed below, both of these types of methods have significant limitations. Our goal is to detect specific actions in realisitic videos with cluttered environments. First, we segment input video into space-time volumes. Then, we correlate action templates with the volumes using shape and flow features. We are able to localize events in space-time without the need for background-subtracted videos.
