Hierarchical Long Short-Term Concurrent Memory for Human Interaction Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human interaction recognition in videos is challenging due to complex, evolving dynamics.
• Existing Long Short-Term Memory (LSTM) methods often fail to capture inter-related dynamics among multiple individuals.
• Current approaches inadequately model the temporal changes in how people interact.

Purpose of the Study:

• To propose a novel method, Hierarchical Long Short-Term Concurrent Memory (H-LSTCM), for recognizing human interactions.
• To effectively model long-term, inter-related dynamics among multiple persons in video data.
• To overcome limitations of existing methods in capturing the nuances of group interactions.

Main Methods:

• Developed a Hierarchical Long Short-Term Concurrent Memory (H-LSTCM) model.
• Utilized Single-Person LSTM to model individual dynamics.
• Introduced a Concurrent LSTM (Co-LSTM) unit with sub-memory units, a cell gate, and a co-memory cell to integrate inter-related motion information.

Main Results:

• The proposed H-LSTCM effectively models long-term inter-related dynamics among interacting persons.
• Experimental results on public datasets demonstrate the superiority of H-LSTCM over baseline and state-of-the-art methods.
• The model successfully captures complex temporal relationships in human interactions.

Conclusions:

• H-LSTCM provides a significant advancement in human interaction recognition.
• The novel architecture effectively addresses the challenge of modeling multi-person dynamics.
• This approach enhances the understanding of complex human behaviors in video analysis.