Manuscript received February 14, 2026; revised March 27, 2026; accepted May 19, 2026; published October 19, 2026.

Abstract—Visual impairment severely restricts independent mobility in crowded urban environments, where unpredictable obstacles such as pedestrians, vehicles, and street furniture pose constant safety risks. The white cane, the most common assistive tool, detects objects only on contact and provides no contextual information. Although computer vision-based obstacle detection offers a promising solution, real-time deployment on resource-constrained Internet of Things (IoT) and mobile devices remains unresolved. Existing approaches are limited by heavy architectures unsuitable for edge deployment and by applying backbone networks, feature fusion, attention mechanisms, and refinement modules independently rather than as a unified system. This study proposes a lightweight end-to-end obstacle recognition framework integrating EfficientNetV2-B0, an adaptive Feature Pyramid Network (FPN) with learnable fusion weights, a Squeeze-and-Excitation (SE) attention module, and a residual refinement block into a single architecture. The key novelty lies in the adaptive scale-aware fusion strategy that dynamically balances multi-scale feature contributions via SoftMax-normalized importance weights—unlike conventional fixed elementwise FPN operations. Evaluated on ten obstacle classes (3133 training, 432 validation images), the proposed model achieves 98.84% accuracy with weighted precision, recall, and F1-Score of 98.92%, 98.84%, and 98.85%, respectively—an improvement of 75.51% over a baseline EfficientNetB0+SE model (23.33%), confirming the effectiveness of integrated architecture for real-time assistive navigation.

Keywords—obstacle detection, visually impaired navigation, EfficientNetV2-B0, Feature Pyramid Network (FPN), lightweight deep learning
 

Cite: Hadi Santoso, Vemas A. Pratama, Mudrik Alaydrus, Misni, and Isack Farady, "An Efficient End-to-End Obstacle Recognition Framework for Assistive Navigation Based on Enhanced EfficientNetV2-B0 and Feature Pyramid Network," Journal of Image and Graphics, Vol. 14, No. 5, pp. 830-840, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).