Predicting Turn Maneuvers of Cyclists Using Bicycle-Mounted IMU with CNN-LSTM

Cycling is a popular and sustainable mode of transportation, but it poses safety risks due to potential collisions with other road users. Predicting turn maneuvers of cyclists and sharing them to the surrounding traffic is crucial to prevent such accidents. However, existing methods for predicting cycling maneuvers rely on external sensors, which are intrusive or unreliable. In this paper, we use a bike-mounted Inertial Measurement Unit (IMU) to detect pre-maneuver indicators like counter steering and tilting; and optimize a Convolutional Long Short-Term Memory Neural Network (CNN-LSTM) model to predict and classify left turns, right turns, and cruising (cycling straight). We evaluate our method by collecting data from controlled cycling scenarios. The results indicate that as the time gap between prediction and occurrence of maneuver gets closer, the accuracy of the model increases, e.g., our model achieves an F1-score of 0.72 when predicting maneuvers 0.5 seconds ahead, and 0.92 when predicting maneuvers 0.25 seconds ahead. Our method is helpful to alert the cyclists and the nearby vehicles of the upcoming maneuvers using back-lights, for example.