Keyframe extraction using linear rotation invariant coordinates

Today, with the improvements in the processing power of video cards, SOC hardware, and smartphones, the use of 3D motion data has expanded considerably beyond video games. At the same time, through these developments, the use of computer animation also increased along with the rapid progress in areas such as augmented reality, virtual reality, and video editing software. Keyframe extraction is a widely applied remedy for issues faced with 3D motion capture based computer animation. In this work, we propose a novel keyframe extraction method. In this method, firstly the skeletal motion is represented in linear rotation invariant (LRI) coordinates. This representation creates a mesh with joint positions of the related frame in the skeletal motion and then applies the transformation of the LRI coordinates. Afterwards, by performing dimension reduction using PCA, the dimensions covering 95% of the data are automatically selected and the summary data is thus acquired. Then, by applying K-means classification, the summary data is divided into clusters and a keyframe is extracted from each cluster using the cosine similarity measure. To validate the results of our proposed method, we conducted an online user study. The results of the study show that 45% of the participants preferred the keyframes extracted using our LRI-based method, surpassing the alternative by 6%.