EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Synergistic Pixel- and Object-Level Anomaly Detection for Robust Autonomous Driving

Chuan-Yuan Huang, Yu-An Chen, Cheng-Tsung Chen, Kuan-Wen Chen

IEEE Access · 2026 · 14 · 85682-85694

DOI: 10.1109/access.2026.3700157 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Chuan-Yuan Huang und 3 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Science and Technology Council, Taiwan,, Industrial Technology Research Institute, Taiwan.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Detecting unexpected obstacles is a critical safety requirement for autonomous driving. Pixel-level anomaly segmentation enables fine-grained localization but often leads to excessive false alarms, whereas object-level detection yields practical bounding boxes but may lack sufficient spatial precision. In practice, false positives are frequently triggered by visually complex backgrounds (e.g., foliage, shadows, and rough textures), particularly in off-road regions, which can lead to spurious warnings and conservative driving behaviors. To address the aforementioned limitations, this study developed a framework called Overestimation-Refined Anomalous Object Detection (OR-AOD), which integrates pixel- and object-level anomaly cues through a bidirectional refinement loop. In this framework, refined pixel-level anomalies guide object attention, and object predictions suppress residual texture-induced false positives in the final anomaly map. The OR-AOD framework uses an overestimation map based on the negative LogSumExp function to enhance sensitivity and employs a lightweight refinement module to regularize its background activation. This framework outputs both a refined anomaly map and anomalous object bounding boxes, providing complementary dense and instance-level cues for downstream decision-making. The proposed framework was tested using the LaF NoKnown and ObstacleTrack benchmarks, and it achieved the Area Under Precision-Recall Curve (AUPRC) values of 86.9% (with a false positive rate, FPR95, of 0.9%) and 90.2%, respectively, supporting reliable hazard perception with reduced false alarms for safety-critical deployment.

Förderung

  • National Science and Technology Council, Taiwan,
  • Industrial Technology Research Institute, Taiwan

Zitieren

APA

Huang, C., Chen, Y., Chen, C., & Chen, K. (2026). Synergistic Pixel- and Object-Level Anomaly Detection for Robust Autonomous Driving. IEEE Access, 14, 85682-85694. https://doi.org/10.1109/access.2026.3700157

Vancouver

Huang C, Chen Y, Chen C, Chen K. Synergistic Pixel- and Object-Level Anomaly Detection for Robust Autonomous Driving. IEEE Access. 2026;14:85682-85694. doi:10.1109/access.2026.3700157

BibTeX
@article{Huang2026,
  author = {Huang, Chuan-Yuan and Chen, Yu-An and Chen, Cheng-Tsung and Chen, Kuan-Wen},
  title = {Synergistic Pixel- and Object-Level Anomaly Detection for Robust Autonomous Driving},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {85682--85694},
  doi = {10.1109/access.2026.3700157}
}
Kostenlos starten →