Objective As a key device in urban rail transit train surveillance systems, the imaging distortion of line scan cameras affects the detection accuracy and efficiency to some extent. Therefore, a feature matching-based image correction method is proposed to improve image quality and detection accuracy, providing a guarantee for train operational safety.

Method The LoFTR algorithm based on Transformer local feature matching is adopted for initial feature point matching, and a multi-level mismatch elimination strategy is designed. The quality and robustness of matching points are improved through a multi-level mismatch filtering mechanism, including horizontal order constraints, neighborhood consistency analysis, and clustering optimization. A piecewise smooth fitting model is used to capture the nonlinear deformation of the images, enabling pixel-level correction by generating a correction matrix. Three existing methods are selected for comparison with the proposed image correction method, and the test data covers 200 actual collected images.

Result & Conclusion  The RMSE (root mean square error) of the corrected image is 0.187, with a structural overlap of 94.9%, and an average processing time of 42.1 ms. Compared with various existing methods, relevant indicators of the proposed image correction method have certain advantages and can meet the accuracy and efficiency requirements of practical applications.