Human action recognition using differential motion

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
Analyzing video and getting information from it is a growing field in computer-vision. The differential feature of a video compared to an image is motion. A video is a very high dimensional data and contains information of the environment, which is important for applications such as navigation, surveillance and video indexing. Therefore, extracting a compact representation of a video is required which can be used for various applications. There are various methods for estimating the motion from a video. One of the popular methods for estimating motion is optical flow. The main theme of this work is to separate motions of different objects, such as background and foreground, by computing higher order derivatives of the motion vector. We proposed a method to compute differential motion on optical flow, which captures the motion of moving objects with respect to their potentially non-stationary backgrounds. Two methods based on curl and divergence for computing differential motion are proposed. The curl of optical flow is used to determine the interest points in video and extract features around it. We demonstrate the robustness of the proposed curl-based detector under common video transformations. We also proposed a descriptor that captures the location of the interest point with respect to neighboring interest points, which contains important information that state-of-the-art descriptors do not capture. The combination of the proposed detector and descriptor in a bag-of-features action recognition framework tested competitively with state-of-the-art methods.
Supervisors: Amit Sethi and Srinivasan Krishnaswamy