A Comprehensive Methodological Survey of Human Activity Recognition Across Diverse Data Modalities

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Human Activity Recognition (HAR) systems are crucial for understanding human behavior.
• HAR utilizes diverse data modalities like RGB, skeleton, audio, and radar signals.
• Significant research exists on various HAR approaches across different modalities.

Purpose of the Study:

• To provide a comprehensive survey of recent advancements in HAR from 2014 to 2025.
• To focus on Machine Learning (ML) and Deep Learning (DL) techniques.
• To categorize approaches based on input data modalities.

Main Methods:

• Review of peer-reviewed research papers in English.
• Categorization of HAR methods by single-modality and multi-modality techniques.
• Analysis of fusion-based and co-learning frameworks, hand-crafted features, human-object interaction recognition, and activity detection.

Main Results:

• Detailed dataset descriptions for each modality.
• Summary of the latest HAR systems and their comparative results on benchmark datasets.
• Mathematical derivations for evaluating deep learning models for each modality.

Conclusions:

• Identification of trends and challenges in current HAR research.
• Proposing effective future research directions in the field of HAR.
• Highlighting the importance of multi-modality and advanced DL techniques for improved HAR performance.