DANGER DETECTION FOR CYCLISTS WITH MACHINE LEARNING (IN THE CITY OF COPENHAGEN).

This paper offers several ways to classify time series data recorded by cyclists in an urban area like Copenhagen to predict and classify dangerous situations and areas. Therefore, several neural networks used a training dataset of bicycle trips consisting of position data and associated system modes derived from a Support Vector Machine. The system modes indicate if cyclists are in dangerous situations. The model used position data and derived features like velocity, acceleration, angular deviation, and the deviation of the previous cycling behaviour in the respective trip. A gated recurrent neural network model achieved the best resulting accuracy of 83 % in a binary classification between accident and no danger. Through this, it was possible to determine if a bicycle accident happened due to the cyclist's environment e.g., cobblestones, or due to their cycling behaviour. This way the dataset and the approved machine learning model can show municipality of cities which spots are currently posing a threat for cyclists. Furthermore, the developed algorithm can pose as a basis for a cyclist app that warns its user about dangerous driving behaviour or upcoming danger spots. All the developed algorithms can be transformed to other cities. [ABSTRACT FROM AUTHOR]