Video Recognition of Human Fall Based on Spatiotemporal Features. Issue 2 (2nd April 2016)

• Record Type: Journal Article
• Authors: Wang, Kai; Zhao, Youjin; Xiong, Qingyu; Shen, Xiling; Fan, Min; Gao, Min
• Is Part Of: Intelligent automation & soft computing. Volume 22:Issue 2(2016)
• Page Start: 303
• Page End: 309
• Publication Date: 2016-04-02
• Subjects: Video recognition -- human fall -- spatiotemporal features -- PCA -- gaussian RBF kernel; Artificial intelligence -- Periodicals; Intelligent control systems -- Periodicals; 003.5
• DOI: 10.1080/10798587.2015.1095487
• Languages: English
• ISSNs: 1079-8587

Abstract

Abstract: A systematic framework for recognizing human fall from video is presented in this work. For the foreground extraction, instead of remodeling background of every video frame, we directly extract cuboids that are composed of spatiotemporal interest points detected by separable linear filter from video sequences. We then represent these video patches as local image gradient descriptors with greatly reduced dimensions by principle component analysis (PCA). From labeled video patches, a supervised learning method based on Gaussian RBF kernel is proposed to determine the maximum margin between fall and normal activity, and then a novel video sequence can be categorize into fall or normal activity by an optimal hyperplane. We tested the above method on datasets set up based on the LPO-CV testing paradigm, which verified the proposed method and demonstrated its advantage over other state-of-the-art approaches for fall recognition.