Description
Riding a bicycle in shared traffic alongside motor vehicles causes discomfort or even stress for many cyclists. Avoiding busy or crowded roads is only possible with good local knowledge, as no data is available on the frequency of encounters with motor vehicles for most roads. Acquiring a data set that combines smartphone sensor data with known vehicle encounters can become the foundation for a smartphone based moving vehicle detector. Therefore, readings from the omnipresent smartphone sensors magnetometer and barometer can be exploited as indicators of passing vehicles.
In this paper, a novel approach is presented to detect vehicle encounters in smartphone sensor data. For this purpose, a modular mobile sensor platform is first constructed and set up to collect smartphone, camera and ultrasonic sensor data in real traffic scenarios. The platform is designed to be used with various sensor configurations to serve a broader set of use cases in the future. In the presented use case, the platform is constructed to create a reference data set of vehicle encounters consisting of location information, direction, distance, speed and further metadata. To this end, a methodology is presented to process the collected camera images and ultrasonic distance data.
Furthermore, two smartphones are used to collect raw data from their magnetometer and barometric sensor. Based on both, the reference and the smartphones' data set, a classifier for the detection of vehicle encounters is then trained to operate on pure smartphone sensor data. Experiments on real data show that a Random Forest classifier can be successfully applied to recorded smartphone sensor data. The results prove that the presented approach is able to detect overtaking vehicle encounters with a F1-score of 71.0\,\%, which is sufficient to rank different cycling routes by their 'stress factor'.
