Skip to content
Conference Open access

Explainable Deep Learning for Vehicle and Pedestrian Detection in Drone Imagery

Sep 2026 · 2026 11th International Conference on Computer Science and Engineering (UBMK) · pp. 402-407 · 0 citations · 19 references

Abstract

This study presents a comparative and explainable deep learning framework for detecting vehicles and pedestrians in drone-based aerial imagery. The VisDrone dataset was used to address key challenges associated with aerial object detection, including small object size, varying camera angles, crowded scenes, and class imbalance. Five object detection models, YOLOv5l, YOLOv8l, YOLO11l, YOLO26l, and Faster R-CNN, were trained under the same experimental conditions using the official VisDrone-DET2019 split. Their performance was then evaluated on a held-out test-dev subset. The models were evaluated using several metrics, including mAP@50, mAP@75, mAP@50:95, AP-small, AP-medium, AP-large, precision, recall, F1-score, and FPS. The results showed that YOLO26l achieved the highest mAP@50 (0.4350), mAP@50:95 (0.2558), and AP-small (0.1543) scores among the evaluated models. YOLOv8l, on the other hand, achieved the highest inference speed at 46.98 FPS. To examine the computational cost of explainability, a Grad-CAM-based analysis was carried out for all evaluated architectures. The findings show that YOLO-based models offer a practical balance between accuracy and inference speed for UAV-based vehicle and pedestrian detection. The Grad-CAM analysis also provides a measure of the additional computational cost introduced by explainability.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.