Accurate 3D Depth Mapping for Robotic Vision Using Fused CNN and Edge-Preserving Cost Aggregation in Stereo Matching

Stereo matching remains a fundamental challenge in computer vision. While deep learning-based methods have demonstrated superior accuracy, their computational demands limit their practicality. Conversely, traditional stereo algorithms offer efficiency but often fail to handle complex textures, occlusions, and illumination variations effectively. To address these limitations, this paper presents a hybrid framework that combines learned and handcrafted techniques for disparity estimation. This method leverages a CNN for discriminative feature extraction, which is then integrated with an enhanced Census Transform to compute a matching cost with improved invariance to radiometric variations. An Edge-Preserving Cost Aggregation (EPCA) module is introduced, which dynamically adjusts support regions based on local geometric structures to maintain depth discontinuities while suppressing noise. Post-Processing is done using a left-right check to further boost accuracy. Comprehensive evaluations on the Middlebury v3 benchmark demonstrate the framework's effectiveness, achieving competitive error rates of 8.48% (all pixels) and 5.03% (non-occluded regions). The proposed method performs well in textured regions while maintaining computational efficiency comparable to several other published methods. Results suggest that carefully designed hybrid systems can bridge the performance gap between learning and traditional methods, offering a balanced trade-off between accuracy and practicality.