Dual-Attention Framework for Enhanced Road Obstacle Segmentation and Anomaly Detection

Autonomous driving systems rely heavily on semantic segmentation to understand road environments; however, existing methods often fail to detect unseen road obstacles that are absent from training datasets, limiting their use in safety-critical applications. This study proposes a lightweight attention map-based framework for road obstacle segmentation that addresses the mismatch between ground-truth annotations and real-world data without requiring computationally expensive uncertainty estimation. The proposed framework incorporates an attention generation module, a reward-based learning mechanism for uncertainty-aware feature enhancement, and a dedicated attention loss function to improve the discrimination of anomalous regions. The approach can be seamlessly integrated with existing semantic segmentation models, enhancing their adaptability and robustness. Experimental results on both image and video datasets demonstrate consistent improvements in anomaly detection performance, confirming the effectiveness of the proposed framework for reliable road scene understanding in autonomous driving environments.

@article{G2026,
  title     = {Dual-Attention Framework for Enhanced Road Obstacle Segmentation and Anomaly Detection},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Priyanka G and Silvia Gaftandzhieva and Mariya Zhekova},
  doi       = {10.14569/IJACSA.2026.0170752},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170752}
}