| Resumo: | Nowadays smartphones are carrying more and more sensors among which are inertial sensors. These devices provide information about the movement and forces acting on the device, but they can also provide information about the movement of the user. Step detection is at the core of many smartphone applications such as indoor location, virtual reality, health and activity monitoring, and some of these require high levels of precision. Current state of the art step detection methods rely heavily in the prediction of the movements performed by the user and the smartphone or on methods of activity recognition for parameter tuning. These methods are limited by the number of situations the researchers can predict and do not consider false positive situations which occur in daily living such as jumps or stationary movements, which in turn will contribute to lower performances. In this thesis, a novel unconstrained smartphone step detection method is proposed using Convolutional Neural Networks. The model utilizes the data from the accelerometer and gyroscope of the smartphone for step detection. For the training of the model, a data set containing step and false step situations was built with a total of 4 smartphone placements, 5 step activities and 2 false step activities. The model was tested using the data from a volunteer which it has not previously seen. The proposed model achieved an overall recall of 89.87% and an overall precision of 87.90%, while being able to distinguish step and non-step situations. The model also revealed little difference between the performance in different smartphone placements, indicating a strong capability towards unconstrained use. The proposed solution demonstrates more versatility than state of the art alternatives, by presenting comparable results without the need of parameter tuning or adjustments for the smartphone use case, potentially allowing for better performances in free living scenarios.
|
|---|