Multi-scale geometric transformer for tire fragment stitching with small feature differences

Abstract Tire is a vital component of any vehicle, and the surface characteristics have a considerable impact on overall performance. Given the limitations of the sensor’s scanning angle, it is imperative that the fragments of the tire be stitched together to reconstruct a complete tire model. Notably, the smoothness of the tire surface’s geometric structure and the subtle differences in small-scale features have resulted in inaccurate stitching and mismatches. This paper proposes an approach based on a multi-scale geometric transformer (MSGT) for high accuracy and robustness. Firstly, the multiscale geometric structure embedding module extracts the details and overall information of the tire surface at different scales. A gating mechanism is introduced to fuse the distance and angle features to enhance the sensitivity and expressiveness of MSGT to small-scale features. Then, we employ a global attention module (GAM) that combines channel and spatial information to enhance feature differentiation, causing the model to select effective features in various regions of the tire. To further tackle mismatch and uneven distribution existing in the tire point cloud data, we propose an adaptive dynamic thresholding network that dynamically adjusts thresholds based on the distribution characteristics in tread and non-tread areas, effectively filtering out low-confidence matched point pairs. The experiments are conducted on the 3DMatch and self-built tire datasets, with the results demonstrating that MSGT significantly outperforms the other four mainstream deep learning methods in terms of stitching accuracy and robustness. Consequently, the superior performance of MSGT in tire stitching scenarios is validated.