- The paper presents an innovative integration of Semi-Global Matching and superpixel segmentation to refine disparity maps for improved terrain modeling.
- The methodology uses SGM for initial disparity estimation and applies SLIC-based superpixel refinement followed by WLS filtering to handle occlusions.
- Experiments on Mars-analog and benchmark datasets demonstrate enhanced 3D reconstruction accuracy and reduced disparity errors.
Stereovision Image Processing for Planetary Navigation Maps with Semi-Global Matching and Superpixel Segmentation (2509.05645)
This paper presents an approach designed to enhance the accuracy and reliability of terrain models for Mars exploration rovers through an innovative stereovision image processing technique. The integration of Semi-Global Matching (SGM) and superpixel segmentation addresses typical challenges encountered in extraterrestrial environments, including low texture, occlusion, and repetitive surface patterns. By leveraging advanced disparity refinement techniques, this research aims to optimize 3D scene reconstruction crucial for the safe navigation of Mars rovers.
Introduction
Accurate terrain mapping is critical for the navigation of rovers on Mars, where unpredictable terrain conditions pose significant risks to exploration missions. Stereoscopic vision offers the potential to generate detailed 3D terrain models from images captured by rover cameras through the process of stereo vision. Traditional stereo matching techniques used in Martian exploration have relied heavily on local block-matching methods, which can be inadequate due to their limited context awareness and susceptibility to noise and occlusion. As an alternative, Semi-Global Matching (SGM) provides a promising balance by performing 1D scanline optimization, thus reducing computational requirements while maintaining accuracy [6].
Superpixel segmentation approaches have also been explored to enhance disparity estimation. Unlike pixel-level processes, superpixels create perceptually significant regions that conform more closely to natural object boundaries, aiding in more refined and context-aware image processing. This paper presents an enhanced disparity estimation technique by integrating SGM with superpixel-based refinement to create more reliable terrain models for autonomous rover navigation on planetary surfaces like Mars.
Terrain Modelling Pipeline
The paper outlines a comprehensive terrain modelling pipeline implemented through stereo vision. The process begins with the acquisition of rectified stereo image pairs from the cameras attached to a rover. Subsequently, pixel-wise disparities are computed to produce a dense disparity map. The disparities are then reprojected into 3D space using the camera's intrinsic and extrinsic parameters, resulting in the creation of a point cloud that characterizes terrain features spatially. Following this, the 3D points are voxelised for efficient terrain analysis and used to generate a 2.5D digital elevation model (DEM), which is ultimately converted into a 2D binary navigation map.
Proposed Disparity Estimation Method
The proposed method focuses on refining disparity maps initially generated via SGM by augmenting them with superpixel-based segmentation. It consists of three primary stages:
- Initial Disparity Map: Using SGM, an initial dense disparity map with subpixel accuracy is obtained by aggregating matching costs along several scanline paths. The energy function used penalizes disparity changes, encouraging smooth surfaces while acknowledging necessary discontinuities.
- Superpixel-Based Refinement: The superpixel-based refinement process incorporates Simple Linear Iterative Clustering (SLIC) to segment the image into perceptually uniform segments called superpixels. Within each superpixel, the disparities are corrected through RANSAC-based plane fitting.
- Post-Processing and Occlusion Handling: An additional refinement is applied to handle occlusions and reduce noise using Weighted Least Squares (WLS) filtering to smooth out irregularities in the disparity map.
The resulting disparity maps demonstrate improved alignment with object boundaries, reduced noise, and more accurate surface detail preservation. Notably, issues such as block artefacts and disparity discontinuities are considerably minimized with this approach.
Experimental Results and Evaluation
The proposed SGM and superpixel-based methodology was evaluated using three datasets, including a Mars-analogue dataset and two benchmark datasets. Upon evaluation, the approach demonstrated improved consistency and terrain structure fidelity in settings with slope variances and occlusion challenges. In particular, large gaps in disparity maps, frequently observed behind rocks, were successfully minimized. Surface details such as minor rock formations and edges were accurately captured.
The evaluation on the Middlebury stereo benchmark demonstrated that this method is competitive, providing a decrease in the percentage of erroneous disparity values across non-occluded regions and full image assessments. The effectiveness in Martian scenarios was confirmed by producing smooth disparity estimates that accurately delineate soil and rock regions, with a resolution of 0.05 meters per pixel and appropriate handling of occlusions using WLS filtering.
Conclusion
The research introduces an effective disparity refinement technique for planetary rover navigation, integrating SGM with superpixel-based segmentation to significantly advance the precision of depth maps. This enhances the rover's ability to make informed navigational decisions based on terrain model constructions. Future work could consider extending the method's applicability to a broader spectrum of extraterrestrial exploration settings or incorporating machine learning techniques to further enhance real-time performance and adaptability.
