MSTCN: A multiscale temporal convolutional network for user independent human activity recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial in healthcare and sports, with existing methods like handcrafted features (HCF) and deep learning (DL) having limitations.
• HCF methods involve complex preprocessing and manual feature extraction, leading to bias and pattern loss.
• Deep learning models like CNNs capture spatial features, while RNNs capture temporal features but suffer from memory issues; LSTMs have longer dependencies but higher computational costs.

Purpose of the Study:

• To introduce a novel Multiscale Temporal Convolutional Network (MSTCN) for enhanced human activity recognition.
• To overcome the limitations of existing HCF and DL methods by reducing preprocessing and manual feature engineering.
• To improve temporal feature extraction and model efficiency in HAR systems.

Main Methods:

• The proposed MSTCN utilizes a temporal convolutional architecture inspired by the Inception model.
• It employs multiscale feature extraction through separable convolutions with varying kernel sizes and dilated convolutions to expand receptive fields.
• Residual connections are incorporated to prevent information loss and mitigate gradient vanishing, enabling longer effective history with low computation.

Main Results:

• MSTCN was evaluated on the UCI and WISDM datasets using a subject-independent protocol.
• The model achieved high accuracies of 97.42% on the UCI dataset and 96.09% on the WISDM dataset.
• These results demonstrate the effectiveness of MSTCN in accurately recognizing human activities.

Conclusions:

• The proposed MSTCN outperforms state-of-the-art methods in human activity recognition.
• MSTCN achieves superior accuracy without the need for manual feature engineering.
• The novel architecture offers an efficient and effective solution for HAR applications.