Analysis of Smartphone Sensor Bias from an Activity Recognition Experiment in the Wild

A multitude of sensor models is being embedded in a variety of smartphone brands. Notably, the accelerometer sensor has been widely used to recognize a variety of human-induced activities leading to the growth of human activity recognition (HAR) research. However, diversities in sensor models create heterogeneity in the data which could result in inaccurate recognition accuracies. Existing empirical work handling a diverse set of mobile devices do not pay attention to this issue. In this paper, we report on the bias observed across 13 different accelerometer models spanning 26 different phones. A continuous sensing application was deployed on all phones over a period of 45 days to empirically assess students’ behavioral patterns (physical activity patterns). Sensor bias analysis revealed One Plus 6, Poco F1 to have the lowest bias and Samsung S4 the largest bias. Except for Samsung S4(11%), Honor 7(6%) and Nexus 5(6%), all other phones showed less than 3% avg deviation from the actual value which was 9.8m/s
²
. Using the bias ranges observed, we classify the entire unlabeled data with accuracy close to 98%. We believe our results will provide a baseline information before using the phones for continuous sensing applications in general and human activity monitoring in particular.