A position-independent method for soil types recognition using inertial data from a wearable device

This paper describes a novel method for recognizing different soil types based on inertial data generated by a user's gait. To achieve this objective, a new wearable device which aims at collecting data produced by an embedded 6-axis accelerometer/gyroscope was designed first. To command this piece of hardware (start and stop recording, as well as annotate raw data), a mobile application was specifically developed. A total of 70 well-known features both from time and frequency domains that are mostly used in activity recognition were computed over each signal to produce enough discriminating characteristics. Then, two machine learning algorithms (Random Forest and k-Nearest Neighbors) were employed to classify such data. The proposed method was tested with 9 participants on four soil types with an experimental setup close to real use case situations. Results obtained let us state that a soil-types recognition is not only possible but also accurate and reliable since overall median F-Score measures of 82% and 86% were obtained respectively with the Random Forest and the k-NN classifiers. Although the user independence of our system was not proven due to a limited number of involved users, the independence condition of the position of the wearable device was clearly demonstrated.