Proposal of a system for detecting hazardous areas in micromobility

This study proposes a system to detect hazardous manholes as road surfaces that pose a fall hazard for micromobility. A method for classifying hazardous points by machine learning from various data from a 6-axis sensor on the electric kickboard was studied. Eleven features were selected as the feature values: the RMS, peak-to-peak, standard deviation, difference in standard deviation between the before and after time windows of the slip and roll angular accelerations and the Crest factor of the slip angular acceleration, which were confirmed to have significant differences in the presence of falls by statistical tests, and the machine learning models were the SVM (Support Vector Machine), effective for binary classification, and the RF (Randum Forest), which are effective for binary classification, were used to construct the model. 11 features were divided into two groups, and models using the logical sum operation were constructed for the outputs of the two machine learning models constructed for each, and the accuracy of each was compared. As a result, the machine learning model with RF using the logical sum operation was able to classify hazardous points with an accuracy of 97% reproducibility.