AER-MFF: Adaptive Edge Refinement with Multi-feature Fusion for Autonomous Driving Semantic Segmentation

This paper presents an advanced post-processing framework to improve the limitations of deep learning-based segmentation models in handling local details and some unexplained segmentation errors. The proposed approach involves the detection of contours from segmentation results, followed by the extraction of local regions surrounding these contours through mapping to the original image. Critical known features are then extracted from these regions, and adaptive weights are calculated based on their reliability across different scenarios to facilitate feature fusion. The optimization process employs an energy function comprising internal energy gradients and fused feature energy gradients. And a new optimization approach, adam with adaptive points removal and movement constraint, is proposed to make the final contour align with the image edges and be smooth, while reducing noise segmentation regions by introducing a distance constraint, thresholds for contour length, and feature gradient variance on contour point updates in Adam. To evaluate the efficacy of the proposed framework, extensive experiments were conducted. When integrated with three types of models, it significantly enhanced segmentation quality. Furthermore, the method exhibited superior enhancement in comparison to other post-processing mechanisms. In real-world tests conducted on an autonomous driving platform, significant performance improvements were observed in campus and open park scenarios, highlighting the robustness and accuracy of the method.