Riding a bicycle is an environmentally friendly and healthy mode of transportation. However, riding a bicycle also comes with the risk of traffic accidents. One approach to reduce accidents is to use cooperative collision avoidance systems that enable smartphones, carried by cyclists, to communicate with intelligent vehicles. Smartphones are not only equipped with GNSS sensors to obtain a cyclist’s position, but they also come with IMU sensors, which can be used to recognise a cyclist’s riding behaviour. Such riding behaviour information can help to improve collision avoidance systems, because the system can adapt its internal trajectory prediction according to the cyclist’s riding behaviour. This results in a more accurate position prediction based on the actual behaviour of the cyclist, and thus a higher chance of avoiding a collision. In this work, we propose our system BikeSense and perform experiments to collect sensor data on different riding behaviours, such as pedalling, coasting, and braking. To classify these riding behaviours, we use a transformer-based classifier. Then, we analyse how different typical carrying positions of the smartphone in body-worn positions (trouser pocket and backpack) and positions for an on-board IMU (handlebar and frame) influence the detection accuracy. Our results show a promising riding behaviour recognition rate with an average f1-score of 96.5%.