<p>With the large-scale application of three-dimensional laser scanning technology in railway inspection vehicles, an increasing number of researchers have utilized laser point cloud data for intelligent detection of railway bridges. Traditional point cloud segmentation methods suffer from segmentation confusion on bridge key components, particularly in regions such as sidewalks and guardrails where locally similar geometric features coexist with significant elevation discrepancies, leading to significant computational errors in eccentric distance calculations. A railway bridge eccentric measurement method is proposed based on dual feature enhancement and structural prior constraints. The dual feature enhancement is implemented through our proposed network - Elevation-enhanced Dynamic Graph Convolutional Neural Network (Elev-DGCNN): (1) The explicit elevation-color encoding at the input level transforms elevation differences into effective color features to distinguish various components; (2) An elevation adaptive weighting module within the network architecture that dynamically emphasizes the importance of elevation features during semantic segmentation. Additionally, to enhance the accuracy of eccentric distance calculations, an Adaptive Iterative Label Correction (AILC) method is introduced to address specific types of residual semantic confusion after initial processing. This approach utilizes spatial distribution constraints of bridge components combined with neighborhood statistics for label correction. Experimental results demonstrate that this method significantly enhances point cloud segmentation accuracy while effectively reducing computational errors in eccentric distance measurements, decreasing the average eccentric calculation error from −34.79 mm to 0.02 mm. Compared with conventional approaches, the proposed method achieves lower measurement errors, with its effectiveness and performance validated through real-world dataset experiments.</p>