Accurate Recognition of the Current Activity in the Presence of Multiple Activities

Sensor based activity recognition (AR) has gained extensive attention in recent years due to the ubiquitous presence of smart devices, such as smartphones and smartwatches. One of the major challenges posed by AR is to reliably recognize the current activity, when a given window of time series data contains several activities. Most of the traditional AR methods assume the entire window corresponds to a single activity, which may cause high error rate in activity recognition. To overcome this challenge, we propose a Weighted Min-max Activity Recognition Model (WMARM), which reliably predicts the current activity by finding an optimal partition of the time series matching the occurred activities. WMARM can handle the time series containing an arbitrary number of activities, without having any prior knowledge about the number of activities. We devise an efficient dynamic programming algorithm that solves WMARM in \(\mathcal {O}(n^2)\) time complexity, where n is the length of the window. Extensive experiments conducted on 5 real datasets demonstrate about 10%–30% improvement on accuracy of WMARM compared to the state-of-the-art methods.