Adaptive sliding window based activity recognition for assisted livings

Abstract The incessant diffusion of smart wearable devices and Body Area Networks (BANs) systems is pushing more and more researches on human activity recognition. Previous studies usually adopted fixed time window or chose an optimal window size depending on the characteristics of the activity. In this paper, we propose an adaptive time window-based algorithm of activity recognition, which addresses the problem of different types of activities having different time duration that usually causes poor recognition results. Multivariate Gaussian Distribution (MGD) method was used to detect the signal difference between each determined time window and the defined activity characteristics, and we more accurately define the time window size for each activity. To evaluate the proposed algorithm, we focused on the activities performed by wheelchair users in their daily life, so as to provide them with better healthcare services. We construct two different datasets, respectively considering static and dynamic wheelchair user activities on different ground surfaces. Then, we use time window expansion and contraction method to determine and dynamically adjust the window size for each activity. Different comparison criteria such as recognition precision and F-score were used to evaluate our algorithm. Experiment results revealed that, according to F-score, the proposed algorithm performs 15.3% better than traditional methods in static conditions. In dynamic scenarios we observed 6.4% and 24.5% improvements on flat and rough floors, respectively.