A Machine Learning Based Video Summarization Framework for Yoga-Posture Video

Video summarization techniques aim to generate a concise but complete synopsis of a video by choosing the most informative frames of the video content without loss of interpretability. Given the abundance of video content and its complex nature, there has always been a huge demand for an effective video summarization technique to analyze various dynamic posture centric videos. Yoga session video summarization is one of the interesting application areas of dynamic posture centric video analysis that is lately drawing the attention of computer vision researchers. The majority of available general video summarizing methods fail to detect key yoga poses in a yoga session video effectively, as they do not consider posture-centric information while extracting key frames. In this paper, we propose a machine learning based video summarization framework, which is capable of extracting a series of key postures in a yoga session video by tracking a few key-posture points corresponding to vital parts of the human body. Compared to the widely used FFMPEG tool, the proposed method appears to have a higher proportion of matched keyframes but a lower proportion of missing key-frames and redundant non key-frames with respect to the ground truth set, demonstrating its potential as an effective yoga posture video summarizer.