Wavelet-based image preprocessing for enhancing trajectory estimation in monocular ORB-SLAM3

Abstract

This paper examines whether classical wavelet-based image preprocessing can improve the accuracy of camera-based simultaneous localisation and mapping (SLAM). In such systems, a moving camera estimates its own trajectory by repeatedly recognising visual structures in consecutive images. The quality of these images therefore, has a direct influence on the stability of localisation, especially when the scene contains noise, weak texture, uneven illumination, or rapidly changing visual details. The applied approach involves preprocessing the input images before they are passed to the localisation algorithm, while keeping the remaining conditions unchanged. Wavelet preprocessing is used because it separates an image into components representing coarse brightness information and finer structural details such as edges, junctions, and textures. This makes it possible to reduce unstable fine-scale fluctuations while preserving image structures that are useful for visual localisation. Five interpretable wavelet-based preprocessing variants are compared with the original monocular image stream using repeated evaluations on standard camera-trajectory sequences with reference motion data. The results show that wavelet preprocessing can improve trajectory estimation when it strengthens stable visual information without distorting the scene structure. The best variant reduces the mean absolute trajectory error by 26.1% compared with the unprocessed monocular baseline. At the same time, the results indicate that the effectiveness of preprocessing depends on how well the transformation preserves information relevant to tracking under demanding motion and scene conditions. The study therefore presents wavelet preprocessing as a low-cost, interpretable image-side support for monocular visual SLAM, together with its practical limitations.