Detection of human activities using multi-layer convolutional neural network.

Area of Science:

• * Human Activity Recognition (HAR)
• * Machine Learning
• * Sensor Data Analysis

Background:

• * Human Activity Recognition (HAR) is crucial for healthcare, sports, and human-computer interaction.
• * Challenges in HAR include achieving high accuracy and robustness with noisy sensor data from accelerometers and gyroscopes.
• * Existing methods often struggle with the complexity of raw sensor data.

Purpose of the Study:

• * To introduce HARCNN, a novel Convolutional Neural Network (CNN) approach for enhanced Human Activity Recognition (HAR).
• * To leverage hierarchical spatial and temporal feature extraction from raw sensor data for improved performance.
• * To demonstrate the robustness and accuracy of HARCNN compared to existing techniques.

Main Methods:

• * Development of HARCNN, a CNN model with 10 convolutional blocks (ConvBlk).
• * Each ConvBlk integrates convolutional layers, ReLU activation, and batch normalization.
• * Fusion of outputs from specific blocks (ConvBlk_3/4, 6/7, 9/10) via depth concatenation, followed by max-pooling.

Main Results:

• * HARCNN achieved high accuracy on multiple datasets: UCI-HAR (97.87%), KU-HAR (99.12%), WISDM (96.58%), and HMDB51 (98.51%).
• * The model demonstrated robustness across various window sizes (50ms to 2s), with optimal performance at 200ms.
• * Significant improvements were noted in minimizing false positives and false negatives compared to traditional CNNs and state-of-the-art methods.

Conclusions:

• * HARCNN effectively extracts hierarchical spatial and temporal features for superior Human Activity Recognition.
• * The proposed model offers a robust and accurate solution for HAR, adaptable to diverse real-world applications.
• * HARCNN's performance highlights its potential for reliable human activity monitoring and interaction systems.