Video Summarization Using Geometric Primitives

Video summarization is the process to extract informative events of a video and represent in the condensed form. The paper proposes a new method for extracting important contents of a video for summarization using geometric primitives, such as line segments, angles, and conic parts. The primitives have the capabilities to represent complex shapes and structures of objects in a video frame. Therefore, they are indeed powerful features to localize objects in complex environments. After localizing objects, a cost function is applied to measure the dissimilarity of locations of geometric primitives to detect the movement of objects between consecutive frames. After detecting the dissimilar geometric primitives, amount of object movements are calculated based on the number and length of each geometric primitives. Utilizing this information, each video frame is assigned a probability score to become a key frame. Finally, a set of key frames are selected as per user preference. The proposed approach is evaluated using BL-7F, Office, and Lobby datasets and obtained better performance than the intra-view results of the recently proposed state of the art GMM based method.