Two-level Ensemble Deep Learning for Traffic Management using Multiple Vehicle Detection in UAV Images

الموضوعات : مهندسی هوشمند برق

Zeinab Ghasemi Darehnaei ¹ (Department of Electrical Engineering, College of Engineering, Islamic Azad University Saveh Branch, Saveh, Iran)
Seyed Mohammad Jalal Rastegar Fatemi ² (Department of Electrical Engineering, College of Engineering, Saveh Branch, Islamic Azad University, Saveh, Iran)
Seyed Mostafa Mirhassani ³ (Department of Electrical Engineering, Islamic Azad University Shahrood Branch, Shahrood, Iran)
Majid Fouladian ⁴ (Department of Electrical Engineering, College of Engineering, Islamic Azad University Saveh Branch, Saveh, Iran)

تاريخ الإرسال : 26 الثلاثاء , ذو القعدة, 1442 تاريخ التأكيد : 10 الأربعاء , محرم, 1443 تاريخ الإصدار : 24 الأربعاء , محرم, 1443

الکلمات المفتاحية: Ensemble Learning, deep transfer learning, multiple object detection, unmanned aerial vehicles,

ملخص المقالة :

Environmental monitoring via vehicle detecting using unmanned aerial vehicle (UAV) images is a challenging task, due to small-size, low-resolution, and large-scale variation of the objects. In this paper, a two-level ensemble deep learning (named 2EDL) based on Faster R-CNN (regional-based convolutional neural network) is introduced for multiple vehicle detection in UAV images. We use three CNN models (VGG16, ResNet50, and GoogLeNet) that have already pre-trained on huge auxiliary data as feature extraction tools, combined with five learning models (KNN, SVM, MLP, C4.5 Decision Tree, and Naïve Bayes), resulting 15 different base learners in two levels. The final class is obtained via a majority vote rule ensemble of these 15 models into five vehicle classes (car, van, truck, bus, trailer) or “no-vehicle”. Simulation results on the AU-AIR dataset of UAV images show the superiority of the proposed 2EDL technique against existing methods, in terms of the total accuracy, and FPR-FNR trade-off.