Hierarchical Sparse Vision Transformers for Real-Time Drone-Based Object Detection
Journal of Artificial Intelligence and Capsule Networks · 26.05.2026 · 8 (2) · 147-165
DOI: 10.36548/jaicn.2026.2.006 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Artificial Intelligence and Capsule Networks, online erschienen am 26.05.2026. Verfasst von Vivekanandam B.. Die Arbeit nennt 25 Literaturangaben.
Abstract (Originalsprache)
Object detectors based on transformers offer high contextual modeling but have a quadratic complexity of attention, which has restricted their application to real-time in aerial settings. The proposal presented in this paper is a Scalable Adaptive Hierarchical Attention Transformer (SAHAT-Det) that is proposed to be effective at detecting objects and objects in drone imagery. The framework presents the concept of dynamic relevance-based token scoring, top K sparse attention calculation, and adaptive token pruning to lower the computational cost. A multi-scale hierarchy fusion module retains small-scale spatial details especially of objects that are small and far away. On the VisDrone dataset, experimental results show a higher mAP of 0.5:0.95 and small-object detection accuracy than the latest CNN and transformer-based baselines that are run under the same configuration. Although there is less attention computation, the suggested model still retains close real-time inference speed. The qualitative analysis also proves enhanced localization stability in high density urbanized scenes. The obtained results show that adaptive sparse attention offers an efficient compromise between the accuracy of detection and processing cost in real-time aerial object detection.
Zitieren
APA
B., V. (2026). Hierarchical Sparse Vision Transformers for Real-Time Drone-Based Object Detection. Journal of Artificial Intelligence and Capsule Networks, 8(2), 147-165. https://doi.org/10.36548/jaicn.2026.2.006
Vancouver
B. V. Hierarchical Sparse Vision Transformers for Real-Time Drone-Based Object Detection. Journal of Artificial Intelligence and Capsule Networks. 2026;8(2):147-165. doi:10.36548/jaicn.2026.2.006
BibTeX
@article{B2026,
author = {B., Vivekanandam},
title = {Hierarchical Sparse Vision Transformers for Real-Time Drone-Based Object Detection},
journal = {Journal of Artificial Intelligence and Capsule Networks},
year = {2026},
volume = {8},
number = {2},
pages = {147--165},
doi = {10.36548/jaicn.2026.2.006}
}
EUVIMED