Real-time detection of rare roadside obstacles using YOLOv8-n in autonomous vehicles.

Tanveer, AB; Kamal, MA; Alam, MM; Su'ud, MM · 2026 · PubMed Central

DOI: 10.1371/journal.pone.0350732

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical safety challenge of detecting rare roadside obstacles—such as traffic cones, fallen trees, debris, barrels, and rocks—in autonomous vehicles (AVs). While standard object detection models perform well on common traffic elements, they often fail to identify these infrequent yet high-consequence hazards, particularly under resource-constrained edge-computing conditions. The authors aim to develop a lightweight, real-time detection framework that maintains high accuracy without requiring expensive hardware, thereby bridging the gap between computational efficiency and safety-critical perception. To achieve this, the researchers constructed a unified dataset by merging multiple open-source repositories, resulting in 6,350 annotated images across five obstacle classes. They selected YOLOv8-n as the base architecture due to its low parameter count and suitability for edge deployment. The model was refined using transfer learning from COCO pre-trained weights and enhanced with a feature-weighted fusion strategy to improve multi-scale contextual integration. To ensure robustness against real-world variability, the training process incorporated data augmentation techniques, including brightness fluctuation, rotation, flipping, and geometric distortion. The model was trained for 300 epochs with a learning rate of 0.001 and a batch size of 16, utilizing early stopping to prevent overfitting. Experimental evaluation on a mid-range NVIDIA P100 GPU demonstrated that the optimized framework achieves significant performance metrics. The model attained a precision of 95.4%, a recall of 93.9%, an F1-score of 94.6%, and a mean average precision (mAP@0.5) of 98.1%. Crucially, the system maintained an inference speed of 68 frames per second, confirming its capability for real-time operation. These results indicate that the lightweight architecture can effectively detect small, occluded, or visually ambiguous rare objects without the computational burden of heavier models. The findings validate that YOLOv8-n, when systematically adapted with targeted augmentation and feature fusion, is highly suitable for edge-based autonomous driving systems. By delivering precise, low-latency detection on resource-limited hardware, the framework offers a viable solution for enhancing AV safety in unpredictable road environments. This work highlights the importance of balancing model complexity with deployment constraints, providing a practical approach for integrating rare-object detection into existing autonomous vehicle perception pipelines.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-25; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• perceptual countermeasures