Predicting Ego Cyclist Trajectories Using Bicycle-Mounted Sensors with CNN-LSTM

Cyclists are vulnerable in traffic, and predicting their future trajectories can enhance safety by informing surrounding road users and enabling timely interventions. However, bicycles pose challenges for accurate and timely predictions due to limited onboard resources. We propose a lightweight CNN-LSTM model for cyclist trajectory prediction using dual-modal input from onboard GPS and inertial measurement units (IMUs). Using our newly collected data from 21 cyclists equipped with GPS and four IMUs mounted on the helmet, handlebar, frame, and pedal, we analyze the impact of sensor placement and modality on prediction accuracy. Our analysis shows that the handlebar IMU yields the best single-sensor performance, while combining all IMUs with GPS achieves the highest accuracy. The model generalizes well to unseen cyclists, achieving 5-second-ahead predictions with a final error of 3.73 m, enabling up to 98% reduction in collision risk. Moreover, the model is computationally efficient for real-time inference, facilitating integration into V2X-based collision avoidance systems for cyclists.