Year: 2026
Type: article
Abstract: Context. Ensuring robust, adaptable, and compute-efficient object detection in UAV aerial imagery under distribution shifts, structured and unstructured noise, and strict onboard latency/energy budgets is an urgent scientific task. A compute-aware detector and a complementary training/adaptation method that integrate a dynamic transformer backbone with gate units, parameter-efficient adapters, and resource-bounded test-time adaptation to sustain accuracy under realistic perturbations and domain shift.Objective. Development of a model and method for object detection in aerial imagery that jointly provide robustness and adaptability while meeting embedded compute and real-time constraints typical of onboard UAV systems.Methods. The approach combines dynamic neural networks with Gumbel-Softmax gate units over a ViT-T/16 backbone, a Simple FPN and a RetinaNet-like one-stage head, budget-aware losses that target a desired dynamic compression rate, structured procedural noise (Perlin, Gabor, Worley) for robustness training, LeakyReLU6 with a straight-through estimator for stable gradients, and test-time adaptation via objectness-weighted marginal-entropy minimization on lightweight adapters.Results. On VEDAI, a gated ViT-T/16 detector reaches [email protected] of 0.77 at ~5.0 GFLOPs and 17.8 FPS, rising to 0.79 with adapters and TTA, whereas a static counterpart attains 0.74 at 9.6 GFLOPs and 10.6 FPS; pretraining with procedural noise lifts accuracy further to 0.80 (gated) and 0.82 (gated+TTA) with minimal compute overhead. Under domain shift (trained on VisDrone, evaluated on VEDAI), dynamic gating and TTA improve mAP from 0.54 to 0.60 without noise pretraining and up to 0.66 with it, sustaining ~5.4–5.6 GFLOPs and ~16–17 FPS within an 8–10 GFLOPs budget on 4×A76 CPUs.Conclusions. The proposed object detection model and method – combining dynamic gating, perturbation-aware training, andbudgeted test-time adaptation – reduce average compute while increasing robustness and adaptability, yielding a superior accuracythroughput trade-off for UAV onboard deployment under real-world disturbances and distribution shifts.
Institution: Sumy State University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
Awards: G5281646582, G798858481, G8531853759
OpenAlex ID: W7166019055