A Novel Human Interaction Recognition via Composite Features and Max Entropy Classifier

Human Interaction Recognition (HIR) is a topic of extensive research, posing a challenge for scholars to understand various human interactions and create reliable systems for their accurate identification in practical contexts. This study aims to address recognition hurdles by introducing robust spatio-temporal techniques for feature extraction that can effectively differentiate between different human interactions. Our proposed HIR system identifies eight complex human interactions from complex ISR UOL 3D dataset, including classes: shaking hands, hugging, help walking, help standing up, fighting, pushing, conversing, and drawing attention. Our methodology involves a composite feature extraction strategy where both full-body and point features are captured using thermal features, LIOP features, kinematic postures, and maximum entropy classifier. The experimental evaluation illustrates the efficacy of our devised approach in real-world scenarios with a recognition accuracy of 86.63 %, 82.4%, 77%, and 88% F1 score, precision, and recall respectively, affirming its suitability for integration into security systems, elderly health care, and violence detection.