On the performance evaluation of object classification models in low altitude aerial data

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object classification in low-altitude Unmanned Aerial Vehicle (UAV) datasets presents unique challenges.
• Evaluating various machine learning (ML) and deep learning (DL) models is crucial for optimizing performance.

Purpose of the Study:

• To comprehensively analyze and compare the classification performance of ML classifiers, handcrafted DL models, and pretrained deep networks on UAV datasets.
• To identify the most effective models for object classification in low-altitude aerial imagery.

Main Methods:

• Implementation of ML classifiers: K-Nearest Neighbor, Decision Trees, Naïve Bayes, Random Forest.
• Development of a handcrafted deep model utilizing convolutional layers.
• Evaluation of pretrained deep learning models: VGG-D, InceptionV3, DenseNet, Inception-ResNetV4, Xception.

Main Results:

• Random Forest achieved 90% accuracy.
• The handcrafted deep model reached 92.48% accuracy, demonstrating superiority over ML classifiers.
• Pretrained VGG16 and VGG19 networks achieved near 100% accuracy.

Conclusions:

• Deep learning models, particularly pretrained networks, offer superior performance for object classification in low-altitude UAV imagery compared to traditional ML methods.
• The study provides a valuable benchmark for selecting appropriate models for aerial image analysis.