*4.3. Field Study*

The aim of the field study dataset was the evaluation of the estimation of the overall distance of a run in an outside and real-world scenario. The *Trajectory* algorithm also worked best for this dataset. With an MAE of 94.0 m, it is comparable to the results of GPS, which also produced an MAE of 82.1 m, and is used in state-of-the-art running platforms tracking athlete performances. Besides, the IMU technology has the advantage that it allows velocity and distance computations indoors or in scenarios where no satellite connection for GPS is available. Based on the presented results, we argue that although the *Trajectory* algorithm has high standard deviations in the lab study for the stride length calculation, these have no major impact on the computation for longer distances based on stride length. We believe this is due to errors canceling out over time. As the subjects' average velocity was 3.48 m/s during the data acquisition, the high velocity range of 5–6 m/s was not reached for the amateur runners that participated in this study. We expect the results to be worse for the high velocity range, which can be reached by professional athletes.

The *Stride time* algorithm showed the worst performance for the field study dataset (MAE of 599.7 m). Despite its best energy efficiency, our results indicate that its accuracy is too low to use for tracking velocity and distance. The *Deep Learning* approach (MAE 194.5 m) performs better than the *Acceleration* approach (MAE 333.1 m). Due to the fact that the the neural network also needs the 6D-IMU data as an input, it has no benefit compared to the *Trajectory* approach, which performs better. The *Acceleration* approach only requires the sampling of the triaxial accelerometer, which makes it more energy efficient. Despite its decreased accuracy, we propose to use this algorithm in use cases where very strict energy limitations occur.
