A Study on Sensor-based Activity Recognition Having Missing Data

Human activity recognition is an important area for various applications. Sensor-based activity recognition deteriorates while partial data are lost. Hence, in this paper, we study activity recognition in the presence of data loss. Earlier, we explored sensor-based activity recognition where we train the data with randomly missed data. It is required to investigate better features for handling missing data. Here, we evaluate activity performance result with missing data environment with various feature combinations for multiple classifiers. Initially, we developed a simulated environment to study the impact of features. Afterward, we evaluated our proposed feature-based method on a benchmark dataset named HASC dataset. The dataset has no missing data. However, to evaluate our approach, we added various levels of missing data randomly and studied the performances. We explored mean, variance, skewness and kurtosis as statistical features based on a time-windowing approach. For classification study, we exploited two classifiers called Naive Bayes and Random Forest. Our approach and study demonstrated satisfactory recognition results under various feature combinations in different situations of missing data.