A Deep Learning Approach to Classify Sitting and Sleep History from Raw Accelerometry Data during Simulated Driving
Georgia A Tuckwell1, James A Keal2, Charlotte C Gupta1
1School of Health, Medical and Applied Sciences, Central Queensland University, Adelaide 5001, Australia.
Sensors (Basel, Switzerland)
|September 9, 2022
Summary
Deep learning models can classify driver sitting and sleep history from thigh-worn accelerometer data. This technology helps identify drivers at risk of fatigue-related impairment, enhancing road safety.
Area of Science:
- Wearable technology
- Machine learning in transportation safety
- Human factors in driving
Background:
- Prolonged sitting and insufficient sleep negatively affect driving performance and safety.
- Objective assessment of driver behavior, including recent physical activity and sleep patterns, is crucial for mitigating risks.
- Current methods for assessing driver fatigue are often subjective or lack real-time data.
Purpose of the Study:
- To apply deep learning algorithms to raw accelerometry data for classifying recent sitting and sleep history in drivers.
- To evaluate the effectiveness of convolutional neural networks (CNNs) in analyzing thigh-worn accelerometer data during simulated driving.
- To determine if objective data can identify drivers at risk of fatigue-related impairment.
Main Methods:
- Eighty-four participants underwent a seven-day laboratory study, wearing a thigh accelerometer during simulated driving tasks.
- Raw accelerometry data were collected during 20-minute simulated drives at two time points daily.
- Two CNN models, ResNet-18 and DixonNet, were trained to classify data into sitting/breaking up sitting and 9-hour/5-hour sleep categories using five-fold cross-validation.
Main Results:
- ResNet-18 achieved higher classification accuracy for both activity (88.6%) and sleep history (88.6%) compared to DixonNet (77.2% and 75.2%, respectively).
- Class activation mapping highlighted distinct movement and postural patterns differentiating between the classified states.
- The study demonstrated significant differences in movement and postural changes between the different activity and sleep history classes.
Conclusions:
- Convolutional neural networks are well-suited for classifying sitting and sleep history using thigh-worn accelerometer data during simulated driving.
- This objective, data-driven approach can identify drivers potentially impaired by fatigue.
- The findings have significant implications for developing advanced driver monitoring systems to enhance road safety.
