The trajectory detection of composite fiber bundles holds significant importance for mechanical performance analysis and process parameter control. However, during the fiber bundle forming process, its surface presents challenges for existing detection techniques due to being difficult to access, having weak textural features, and exhibiting local high reflectivity characteristics. Addressing these issues, this paper proposes a novel measurement method based on 3D vision reconstruction technology to achieve high-precision measurement of fiber bundle trajectories in a non-contact manner. In this method, a "coarse to fine" stereo-matching strategy is introduced. Leveraging the constraints of binocular views, an initial matching strategy tailored for weak textural features is established. Furthermore, based on the locally obtained subarea images from the initial matches, a multi-directional fusion of feature enhancement and local stereo matching using deformation optical flow is proposed, enabling higher precision in 3D visual measurement. Through experimental verification using the composite shell winding process, the proposed method is demonstrated to efficiently and accurately capture the trajectories and winding angles of fiber bundles.