Deep-HAR: an ensemble deep learning model for recognizing the simple, complex, and heterogeneous human activities.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for applications like surveillance, wellness, and healthcare.
• Existing HAR methods often struggle with complex and heterogeneous activities, focusing mainly on simple ones.
• Recognizing diverse human actions remains a significant research challenge.

Purpose of the Study:

• To propose a novel Deep-HAR model for recognizing simple, complex, and heterogeneous human activities.
• To leverage the strengths of Convolutional Neural Networks (CNNs) for feature extraction and Recurrent Neural Networks (RNNs) for pattern recognition in time-series data.
• To evaluate the model's performance on diverse human activities.

Main Methods:

• Developed a hybrid Deep-HAR model by ensembling CNNs and RNNs.
• Utilized CNNs for extracting relevant features from activity data.
• Employed RNNs to identify temporal patterns within the sequential data.
• Validated the model on three public datasets: WISDM, PAMAP2, and KU-HAR.

Main Results:

• The proposed Deep-HAR model demonstrated high performance across all activity types.
• Achieved exceptional accuracy rates: 99.98% for simple activities, 99.64% for complex activities, and 99.98% for heterogeneous activities.
• Experimental results confirm the model's effectiveness in comprehensive human activity recognition.

Conclusions:

• The novel Deep-HAR model successfully addresses the challenge of recognizing simple, complex, and heterogeneous human activities.
• The ensemble of CNNs and RNNs provides a robust approach for HAR, outperforming existing methods.
• This research contributes a highly accurate and versatile solution for advanced human activity recognition.