PROJECT TITLE :

Outlier Detection in Wearable Sensor Data for Human Activity Recognition (HAR) Based on DRNNs

ABSTRACT:

Wearable sensors enable the development of tailored apps by providing a user-friendly and non-intrusive approach for extracting user-related data. Human activity recognition (HAR) plays a significant part in the characterization of the user context in such applications. Outlier detection methods look for unusual data samples that are most likely the result of a separate cause. This study combines outlier detection and HAR by presenting a novel method that can recognize data from secondary activities within the main activity as well as extract data segments of a certain sub-activity from a separate activity. Several Machine Learning techniques based on the study of time sequences generated by wearable sensors have previously been employed in the field of HAR. Previous research has shown that deep Recurrent Neural Networks (DRNNs) are best suited to the sequential properties of wearable sensor data. This paper proposes a DRNN-based approach for outlier detection in HAR. The results are validated using two separate datasets for both intra- and inter-subject situations, as well as outlier detection and sub-activity recognition. The concept is trained and validated using a first dataset of 15 people performing four key activities (walking, running, climbing up, and down). Intra-subject outlier recognition detects all major outliers in this dataset's walking activity, however inter-subject outlier detection fails only for one participant who performs the activity in an unusual way. Finding and extracting walking segments present in the other three activities in this dataset was used to validate sub-activity detection. To test the universality of the results, a second dataset was created with four different individuals, a new setting, and various sensor equipment.