Akıllı saat sensörleri kullanarak sigara içmeyi tanıma

Thanks to the technological advances, the use of smartwatches and other wearable devices is growing rapidly. They are equipped with various motion sensors and this makes them effective devices for human activity recognition. While smartwatches can be used to detect complex activities where hand-wrist movements play an important role, smartphones are more convenient to detect simpler locomotion activities. Moreover, an accelerometer is mostly sufficient to detect simple activities, such as walking, with good performance but a gyroscope can increase the recognition rate of more complex activities, such as smoking while walking. This also holds for other parameters, such as sensor sampling rate, feature set and window size, meaning that different activities require different settings for good identification. Besides, changing the parameters can cause higher and unnecessary resource consumption or vice versa on these resource limited devices. In this study, our main motivation is to explore the parameter space that may affect the recognition process in terms of accuracy and as well as resource usage on a large and complex dataset. For this purpose, we collected a dataset of 45 hours from 11 participants. The dataset includes ten different activities including smoking activities in four different postures, such as smoking while standing, some other activities which involve similar hand-wrist movements, such as drinking and some other simple activities, such as sitting. Firstly, we analyze the impact of parameters using 4 different window sizes and overlaps, 63 different features extracted from each sensor, 4 different sensors, 2 different sensor combinations, 3 classifiers and 10 different activities. By changing the values of the mentioned parameters, in the datasets, we gather the recognition accuracies and find the best parameter sets to maximize the recognition performance. Additionally, we analyze the impact of participants' height on the recognition performance. The results show that simple time-domain features perform the best and while the combination of accelerometer and gyroscope sensors performs better for complex activities and accelerometer alone is sufficient for simple activities. When we consider the impact of height on the recognition performance, the results show that it does not have a significant effect when all activities are considered, however, it does have an effect on smoking while standing, particularly for participants with a significant height difference than the others. Secondly, we investigate context-aware activity recognition where parameters are selected on demand. We propose a dynamic parameter selection algorithm, which activates different sensors, sampling rates, window sizes and features on demand according to the type of the activity (simple or complex). This algorithm gets the type information from a state detection algorithm which identifies whether the user is performing a simple or a complex activity. We evaluate the performance of the algorithm both in terms of recognition rate and resource consumption and compare with using static and semi-dynamic parameters. We use feature sets of our first analysis, then we determine the high impact features in order to reduce the number of features and choose the most efficient ones, by applying feature selection algorithms. Results show that, both before and after feature selection, the dynamic parameter selection algorithm achieves 2 to 13% better recognition rate depending on the activity. Dynamic parameter selection, before applying feature selection, consumes 33% less energy and 20% less CPU time, compared to using static parameter selection. Additionally, using selected features in the dynamic parameter selection algorithm, we observe a decrease of 65% on the CPU and energy consumption over the last improvement.