Contour Feature-Driven Deep Learning Method for Arch Dam Point Cloud Registration

Abstract The point cloud of an arch dam comprises two key target categories: the dam body and the bedrock. The bedrock exhibits complex, variable structures and anisotropic surface textures, presenting significant differences from the artificially constructed arch dam. To fully capture the distinctive features of both target types and enhance the registration accuracy of the arch dam, this paper proposes a deep learning-based point cloud registration method driven by arch dam contour features. This approach first addresses the substantial differences in surface features between dam body and rock mass by proposing a contour feature extraction method based on the transformer framework and collaborative clustering. This method separately acquires the contours of both the dam body and rock mass. Subsequently, to overcome the information loss during dimensionality reduction of high-dimensional features by traditional multilayer perceptron, which leads to estimation errors in the transformation matrix and causes local misalignment or global offset in registration results, this paper calculates 12-dimensional feature vectors from the positional relationships between points as input to the registration network. A channel attention mechanism combining average, max, and random pooling is constructed to extract deeper point cloud features from the arch dam data. Finally, a rigid transformation estimation network computes the transformation matrix. The proposed contour feature extraction method captures richer details for registration. Ablation experiments demonstrate that incorporating the feature extractor reduces rock mass and dam registration errors from 24.08 and 23.9 m to 0.71 and 0.64 m, respectively. Following the introduction of the multipooling attention mechanism, the errors were further reduced to 0.0181 and 0.0169 m, respectively. The proposed method provides a robust data foundation for detailed three-dimensional modeling and high-precision deformation detection.