Human Activity Recognition With Accelerometer and Gyroscope: A Data Fusion Approach

This paper compares the three levels of data fusion with the goal of determining the optimal level of data fusion for multi-sensor human activity data. Using the data processing pipeline, gyroscope and accelerometer data was fused at the sensor-level, feature-level and decision-level. For each level of data fusion four different techniques were used with varying levels of success. This analysis was performed on four human activity publicly-available datasets along with four well-known machine learning classifiers to validate the results. The decision-level fusion (Acc = 0.7443 +/- 0.0850) outperformed the other two levels of fusion in regards to accuracy, sensor level (Acc = 0.5934 +/- 0.1110) and feature level (Acc = 0.6742 +/- 0.0053), but, the processing time and computational power required for training and classification were far greater than practical for a HAR system. However, Kalman filter appear to be the more efficient method, since it exhibited both good accuracy (Acc = 0.7536 +/- 0.1566) and short processing time (time = 61.71ms +/- 63.85); properties that play a large role in real-time applications using wearable devices. The results of this study also serve as baseline information in the HAR literature to compare future methods of data fusion.

Automated summary

This study compared three levels of data fusion and found that decision-level fusion outperformed sensor-level and feature-level fusion in accuracy, but required significantly more processing time and computational power. The Kalman filter emerged as a more efficient method with good accuracy and short processing time, making it suitable for real-time applications using wearable devices. The results also provide baseline information for comparing future methods of data fusion in the literature on human activity recognition.