Bike Ego-Trajectory Mapping Using a CNN-LSTM With IMU and Monocular Camera

This paper presents a CNN-LSTM-based approach for mapping bicycle ego-trajectories based on inertial-camera data, capable of accurately capturing non-straight-line movements such as turns and curves. I introduce a novel bicycle trajectory dataset that integrates recordings from a 6 DOF IMU and two cameras, supplemented by GNSS ground truth data. The study systematically evaluates 25 different CNN-LSTM configurations, varying architectures and input parameters, with the top-performing model achieving a RMSE of 4.4 m on a 134 m trajectory. The best-generalised model achieved a RMSE of 33 m on a 598 m trajectory. Although the models are not directly usable for urban trajectories, they lay the foundation for the development and subsequent adoption of bike mapping algorithms. The code and dataset are open source to facilitate further research in this area.