Keyframes and shot boundaries: The attributes of scene segmentation and classification

Abstract

Video analytics of real-life scenario deals with the multimedia data statistics that may be characterized by multimodal features of the video components. Large varieties of low-scale multimodal features of the objects creates many challenging issues for discrimination and analysis. On the other hand occlusion, varied illuminations, and complex environmental conditions highlight the video parsing, a challenging research problem. For the experimental purpose, the vital components of the videos include scenes, shots, keyframes, objects, and background. In this work, we focus on keyframes and shot boundaries for scene segmentation of the sample videos taken from YouTube. Structure Similarity index (SSIM) of the shots is computed from the histograms of LBP and HSV color similarities. Motion similarity and inverse time proximity are added to generate Shot Similarity Graph. Sliding window methods are used for grouping similar shots. The proposed work for scene segmentation is validated on six videos of various semantics characterized by human being and animals. The play of the video ranges from 0.5 to 15min and total no. of scenes in the videos range from 06 to 33. © Springer Nature Singapore Pte Ltd. 2019.