Human Activity Recognition: Review, Taxonomy and Open Challenges.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for various applications, utilizing vision and sensor data.
• Existing literature reviews on HAR require updating due to rapid advancements.
• This study focuses on HAR research published from 2018 onwards.

Purpose of the Study:

• To provide an updated overview of the current state of Human Activity Recognition literature.
• To classify recent HAR research based on application areas, data sources, techniques, and challenges.
• To identify under-explored areas within HAR research.

Main Methods:

• Systematic review and classification of 95 research articles published since 2018.
• Analysis of application domains, data sources (e.g., CCTV, mobile sensors), and recognition techniques (e.g., CNN, LSTM, SVM).
• Identification and discussion of open research challenges and limitations in HAR.

Main Results:

• Most research focuses on daily living and individual/group activities.
• Limited literature exists for real-time activities like surveillance and healthcare monitoring.
• Closed-Circuit Television (CCTV) and mobile sensor data are prevalent data sources.
• Convolutional Neural Network (CNN), Long short-term memory (LSTM), and Support Vector Machine (SVM) are dominant techniques.

Conclusions:

• HAR research has advanced significantly, with established methods for common activities.
• There is a need for increased focus on real-time HAR applications, particularly in surveillance and healthcare.
• Future research should address identified limitations and emerging challenges in HAR.