Cycling accidents are rising, and most one-sided crashes stem from a loss of control whose cause is rarely understood. The kinematics that would explain such events, such as roll, steering and crank angles in the moments before a crash, are not captured by the, sometimes, available video footage of the crash. This thesis asks whether these kinematics can instead be reconstructed from a single, static, monocular RGB camera.

The core obstacle is data. Unlike human pose estimation, there is no large public dataset of real bicycle motion with three-dimensional ground truth. The work therefore contributes a synthetic data generator that combines MuJoCo rigid-body simulation with photorealistic Blender rendering to produce physically plausible bicycle motion with exact two- and three-dimensional keypoint annotations. Second, it introduces 3D-BiKE, a modular pipeline tool that detects the bicycle with RF-DETR, estimates two-dimensional keypoints with RTMPose, and lifts them to root-relative three-dimensional keypoints with PoseMamba, a neural state-space model, from which roll, steering, and crank angle are extracted. Third, it assesses whether a state-space lifter designed for articulated human motion transfers to the rigid dynamics of a bicycle.

Trained only on synthetic data, the pipeline reaches a mean per-joint position error of 17.6 mm on validation data, better than human pose estimation methods trained on far larger real datasets. The derived roll, steering, and crank angles are recovered to 1.4°, 3.0°, and 5.1° root-mean-square error, substantially improving over the only comparable prior bicycle estimator. These results, though using only synthetic data, establish monocular keypoint lifting trained as a credible foundation for video-based cycling safety analysis, with validation on real footage as the essential next step. ... Read more