Author(s): B Vijaya Durga, K Induteja, K Priyanka and B Rupali

Abstract: In this work, we investigate drone autonomous landing scene detection via knowledge transfer. The challenges associated with aerial remote sensing—namely, the fact that various images have distinct representations at different altitudes or that some pictures are very similar—led us to use a deep convolutional neural network (CNN) that is based on knowledge transfer and fine-tuning to address the issue. Next, the seven classes comprise the LandingScenes-7 dataset is created. Furthermore, we use thresholding in the prediction step to take care of the classifier's ongoing novelty detection issue by excluding additional landing scenes. The adaptive momentum (ADAM) optimization technique is used in conjunction with the ResNeXt-50 backbone to facilitate our transfer learning approach. We also compare momentum stochastic gradient descent (SGD) optimizer with ResNet-50 backbone. ResNeXt-50, which uses the ADAM optimization method, performs better, according to the experiment findings. It is possible for drones to autonomously learn landing scenes using this pre-trained model and fine-tuning, as it achieves 97.8450% top-1 accuracy on the LandingScenes-7 dataset.

DOI: 10.33545/27076571.2024.v5.i2b.103

Pages: 115-118 | Views: 1161 | Downloads: 543

Download Full Article: Click Here