Information Entropy-Driven Sequential Fuzzy Indexed Search Trees for Human Activity Recognition

Real-time activity recognition is an important field of machine learning, which can provide useful information about the people inhabiting the observed space. In this paper, a novel system is proposed that uses a so-called Information Entropy-driven Sequential Fuzzy Indexed Search Trees classifier that realizes fast activity recognition, using pose information from the output of a real-time multi-person keypoint detection framework that detects human bodies in image frames and outputs a set of keypoints. The proposed system implements a flexible expert knowledge-based classification in the form of indexing tables and binary search trees, to significantly reduce the search area of the problem space that needs to be regarded during the recognition phase, and thus, significantly lowering the required operation time. The performance of the method is evaluated on five datasets and compared to the performance of two state-of-the-art classifiers.