In this dissertation, several novel approaches to improve overall urban mobility with bikes are presented and raise awareness for bike-specific research.
During our research, we found a research gap for the bike-specific pervasive/ubiquitous applications, which we addressed in our initial works.
For our first approach, we identify distinct bike types (and e-scooters) using smartphone sensor data.
Bike type identification is essential to provide context for context-aware navigation services that take bike-specific aspects into account.
Aspects such as road conditions to improving safety and comfort by suggesting roads that are suitable for the rider's bike type.
We first compare four machine learning classification methods with each other to find the classifier best suited for the problem.
The tested classification methods are Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Perceptron (simplistic neural network) and Random Forest (RF).
The evaluation showed that the CNN method performed the best.
Based on the initial results we adopt a CNN deep learning approach to discern between different types of bikes.
Furthermore, we demonstrate the applicability of our approach with real world cycling data. The evaluation includes different roads, cyclists, types of bikes and smartphones.
Next, we present an approach for predictive maintenance using acceleration data.
Smart features included in modern bikes allow for early detection of deteriorating brake performance.
This enables individual user feedback based on the current condition of their bike and therefore improves safety.
Here, we evaluate the viability of different machine learning approaches for predictive maintenance using acceleration data from bike trips with good and poor braking performance.
We compare two methods of measuring accelerations, specifically hall sensors and inertial sensors.
Our evaluation shows promising viability results, with inertial sensors being more suitable for measuring acceleration data than hall sensors.