On video-based human action classification by SVM decision tree

Abstract

This paper presents a home-care system for recognizing six kinds of daily activities (including walk, jogging, in-place actions like standing, sitting and squat, stand-to-sit, stand-to-squat, and fall) from videos by a multi-SVM classifier with decision tree structure. The system first detects human blobs by a non-parameter background subtraction method, then extracts shape and motion features from those human blobs to discriminate different activities by a multi-SVM classifier. Shape and motion features to character actions are extracted from the minimum bounding box of human blob and motion energy sequence (MES) defined in this paper. The thought of hierarchical classification is introduced to recognize multiple actions. An SVM decision tree classifier is designed experientially and experimentally. Each SVM on the decision tree is trained and tested separately to achieve its best classification performance by choosing proper features and parameters. Due to lack of publicly available action data set focusing on all the aforementioned daily activities, experimental results upon a home-made data set show the perfect identification performance and the robustness of the system on realistic videos.