The effectiveness of simple heuristic features in sensor orientation and placement problems in human activity...

Area of Science:

• Human Activity Recognition (HAR)
• Sensor Signal Processing
• Machine Learning for Wearable Devices

Background:

• Smartphone sensor data for HAR is hindered by sensor orientation and placement variations.
• Orientation and placement variations alter sensor signals for specific activities.
• Orientation and position invariant features offer a solution by minimizing signal alteration.

Purpose of the Study:

• Evaluate the effectiveness of four simple heuristic features in solving sensor orientation and placement problems in HAR.
• Assess the performance of a 1D-CNN-LSTM model using these heuristic features on a large dataset.

Main Methods:

• Collected accelerometer data from 42 participants performing six daily activities at varying intensities (3, 5, 7 METs).
• Recorded data with smartphones in three positions: pocket, backpack, and hand.
• Extracted simple heuristic features and trained/tested a 1D-CNN-LSTM model.

Main Results:

• Achieved 70-73% accuracy in intra-position evaluations (same sensor position for training and testing).
• Obtained 59-69% accuracy in inter-position evaluations (different sensor positions).
• Heuristic features showed greater effectiveness in recognizing high-intensity activities.

Conclusions:

• Simple heuristic features are effective in mitigating sensor orientation problems in HAR.
• Combining heuristic features with placement-focused methods can improve performance for sensor placement issues.
• Heuristic features demonstrate higher efficacy for recognizing high-intensity activities.