This paper addresses the challenge of multi-scale mixed noise in track point clouds acquired through 3D laser scanning, which significantly impacts subsequent feature extraction and 3D model reconstruction. An improved bilateral filtering-based denoising algorithm is proposed to mitigate this issue. Firstly, isolated noise is removed by density threshold and connectivity analysis. Then, the track point cloud is divided into feature and non-feature areas using a classification method based on the Euclidean norm of the point cloud normal vectors, and improved bilateral filtering and least squares plane fitting are used to smooth and denoise the point clouds in the respective areas. The experimental results show that for two sets of different track point cloud data, the algorithm achieves better results with P2point values of 1.647 3 mm and 0.953 7 mm, and P2plane values of 1.246 7 mm and 0.903 7 mm, respectively. Furthermore, the algorithm effectively retains track feature information, offering a theoretical foundation for the rapid and precise extraction of track information.