Predicting Cyclist Trajectories with BicycleNet – MSc Thesis by Gijs de Smit

We’re excited to highlight the recent MSc thesis defense of Gijs de Smit, supervised by Yanqiu Huang and Akhil Pallamreddy.

Gijs developed BicycleNet, a lightweight deep learning model that predicts how cyclists move up to 5 seconds into the future using data from sensors mounted directly on the bicycle. His work stands out for its real-world applicability—using data from 63 participants—and for offering an efficient solution suitable for on-bike deployment.

Among the key findings:

• The best prediction accuracy came from combining all IMUs (helmet, handlebar, frame, pedal) with GPS.
• The pedal-mounted IMU proved to be the most effective standalone sensor.
• BicycleNet achieved competitive performance with 3.5× fewer parameters than traditional models.

This research offers promising insights for improving cyclist safety through smarter, real-time trajectory prediction.

📄 Thesis: Read it here

Congratulations to Gijs and the supervising team!