A multi-layer fusion image enhancement method for visual odometry under poor visibility scenarios

Abstract

Visual odometry (VO) is important in robotic rescue and navigation operations; however, poor visibility scenarios, which are characterised by weak illumination, few textures, and self-similarity, seriously influence image quality and image matching capability, thus degrading VO performance. This study proposes a multi-layer fusion image enhancement method to ensure VO performance under poor visibility scenarios. Our method can achieve a trade-off between naturalness preservation, texture feature enhancement, contrast improvement, and noise suppression. To achieve this, we first decomposed input images into illumination, reflectance, and detail layers. Subsequently, we designed multiple weighting strategies in the illumination layer to preserve naturalness and improve contrast and applied a linear operation to enhance the colour contrast in the reflectance layer. A double-interval frequency optimisation model was also applied to enhance the textures in the detail layer. Finally, these layers were recombined to obtain an enhanced image. Our experimental results demonstrated our quantitative and qualitative superiority over six other methods under poor-visibility scenarios. The overall performance revealed that our method can effectively improve image matching and VO results.

Keywords

• Multi-layer fusion
• Image enhancement
• Poor visibility scenarios
• Image matching
• Visual odometry