EUVIMEDEUVIMED entdecken
Crossref

SCA-YOLO: an improved YOLO algorithm for road object detection in autonomous driving

Yuhao Song, Peng Ji, Tao Wang

Engineering Research Express · 01.06.2026 · 8 (11) · 115228

DOI: 10.1088/2631-8695/ae76eb ↗

Kurz gefasst

Fachartikel in der Zeitschrift Engineering Research Express, online erschienen am 11.06.2026. Verfasst von Yuhao Song und 2 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Road object detection in autonomous driving faces three coupled challenges: severe feature degradation of distant tiny objects under multi-scale variations, persistent background interference in complex traffic scenes, and the stringent real-time and resource constraints of in-vehicle perception systems. To address these issues, this paper proposes SCA-YOLO, a lightweight real-time detector built upon the YOLOv11 baseline. Guided by the depth-dependent properties of feature maps—the trade-off between spatial resolution and channel dimensionality, and the corresponding shift from geometric priors in shallow layers to abstract semantics in deep layers—we systematically redesign the Backbone and Neck through three coordinated modules. The Spatial-Decoupled Aggregation Block separates spatial feature extraction from channel fusion via depth-wise convolutions to preserve high-frequency geometric priors in the shallow backbone. The Semantic Channel Recalibration Attention-Cross Stage Partial module replaces the original C2PSA at the deepest stage, employing multi-subspace channel recalibration to address deep-stage channel entanglement while avoiding the cost of spatial self-attention. The Scale-Aware Neck breaks the isomorphic fusion paradigm through an asymmetric allocation strategy, in which the Channel Recalibration Aggregation Block maintains fusion purity at the P3 and P4 stages while the lightweight VoVGSCSPC module curtails the cost of the P5 stage. Experiments on the KITTI, custom private, and BDD100K datasets show that SCA-YOLO achieves statistically significant mAP@0.5 improvements of 3.0%, 3.6%, and 4.4% over the YOLOv11s baseline, while reducing the parameter count from 9.41 M to 6.62 M and the computational cost from 21.3 to 14.1 Giga floating-point operations. On-device benchmarking on the NVIDIA Jetson AGX Orin platform integrated into our experimental vehicle further confirms a 27.1% inference speed-up over the baseline at 181.3 FPS, comfortably exceeding the on-board camera frame rate and validating the practical edge-deployment efficiency of the proposed model.

Zitieren

APA

Song, Y., Ji, P., & Wang, T. (2026). SCA-YOLO: an improved YOLO algorithm for road object detection in autonomous driving. Engineering Research Express, 8(11), 115228. https://doi.org/10.1088/2631-8695/ae76eb

Vancouver

Song Y, Ji P, Wang T. SCA-YOLO: an improved YOLO algorithm for road object detection in autonomous driving. Engineering Research Express. 2026;8(11):115228. doi:10.1088/2631-8695/ae76eb

BibTeX
@article{Song2026,
  author = {Song, Yuhao and Ji, Peng and Wang, Tao},
  title = {SCA-YOLO: an improved YOLO algorithm for road object detection in autonomous driving},
  journal = {Engineering Research Express},
  year = {2026},
  volume = {8},
  number = {11},
  pages = {115228},
  doi = {10.1088/2631-8695/ae76eb}
}
Kostenlos starten →