On the activity detection with incomplete acceleration data using iterative KNN classifier

In real time continuous activity recognition systems, utilization of a data segmentation stage increases the dependency of success ratio on the size of activity set as well as activity type, duration and sensor sampling rate. In this study, we analyzed if iterative K-Nearest Neighbour based knowledge discovery performed on acceleration data can substitute for the segmentation stage to reduce these dependencies for hand oriented activities. To this end, we compared peak frequency and wavelet entropy feature extraction schemes for the recognition of open-pill-box, put-pill-in-mouth, drink and put-glass-back actions which constitute as a whole “medication intake” activity. We evaluated the performance of these schemes on incomplete data, resulting from the iterative process. According to our findings, peak frequency outperforms wavelet entropy in terms of Intra-Class-Correlation (ICC) metric which is an indication for better adaptation to variation in activity type with incompleteness mitigation problem. Our work points out that a more proficient iterative classification algorithm is required for attaining higher adaptability to the diversification of actions that stem from the incompleteness problem.