BicycleNet : A Temporal Convolutional Network for Ego Cyclist Trajectory Prediction Using Multi-Modal Bicycle-Mounted Sensors

Cycling is a popular and indispensable mode of transportation, but there are many traffic accidents involving cyclists. Predicting cyclist trajectories could prevent accidents by sharing them with surrounding traffic for timely warnings and interventions. This paper introduces two custom lightweight multi-modal models for cyclist trajectory prediction: BicycleNet, a Temporal Convolutional Network (TCN), and a Convolutional Long Short-Term Memory Neural Network (CNN-LSTM) model. The models are evaluated on unseen cyclists and new locations using two large, realistic datasets comprising sensor data from 63 participants, including four inertial measurement units (IMUs) mounted on the helmet, handlebar, frame, and pedal, as well as GPS and a forward-facing camera. BicycleNet achieves an average distance error of 1.86 m and a final distance error of 3.50 m over a prediction horizon of 5 seconds. Both models achieved similar prediction accuracy, while BicycleNet uses 3.5× fewer parameters than the CNN-LSTM. To determine an optimal sensing configuration, an analysis into sensing modality and IMU placement was carried out, which revealed that the best single IMU placement is on the pedal, closely followed by the frame and handlebar, while the helmet performed worst. Combining all four IMUs with GPS provides the best overall performance.