Volume: 19 Issue-01 (January-June) 2024

---

AUTONOMOUS LANDING SCENE RECOGNITION BASED ON HYBRID ENSEMBLE RANDOM FOREST MODEL

ABSTRACT

This study focuses on drones' recognition capabilities of their landing scenes during data transfer. To tackle the problems of aerial remote sensing, including highly similar images or scenes with different representations at different heights, we employ a deep convolutional neural network (CNN) that is based on knowledge transfer and fine-tuning. This results in the creation of the Landing Scenes-7 dataset, which is divided into seven categories. Moreover, we use the thresholding technique to eliminate more landing scenes during the prediction stage, which resolves the classifier's ongoing novelty detection problem. We apply a transfer learning technique using the Hybrid Ensemble Random Forest model and the ResNeXt-50 backbone. We also assess the momentum stochastic gradient descent (SGD) optimizer and the ResNet-50 backbone in conjunction with it. The experimental results show that ResNeXt-50 using the ADAM optimisation approach performs better. With a pre-trained model and some fine-tuning, drones will soon be able to learn landing scenes on their own; this model obtained 97.8450% top-1 accuracy on the LandingScenes-7 dataset.

FULL TEXT

PDF