Autonomous Landing Scene Recognition based on...

Autonomous Landing Scene Recognition based on Transfer Learning for Drone

Author Name : Dr. Regonda Nagaraju, P. Tejaswini, S. Ravindra, N. Renuka, A. Revanth, Sheik Rezwan

ABSTRACT The project was entitled “Autonomous Landing Scene Recognition Based on Transfer Learning for Drones”. This project focuses on improving autonomous landing scene recognition for drones, addressing challenges like similar or varied scene representations from different altitudes. A key difficulty in aerial remote sensing is that many landing zones appear similar or differently depending on the drone's altitude, making it challenging for the system to distinguish safe landing spots. To tackle this, we use a deep convolutional neural network (CNN) with knowledge transfer and fine-tuning. Specifically, we fine-tune a pre-trained model to improve its ability to recognize landing zones in aerial images. To train and evaluate our model, we create a new dataset, LandingScenes-7, which includes images from seven different landing scene categories. An additional challenge in this task is detecting novel or previously unseen landing zones. To address this issue, we introduce a thresholding method during the prediction stage to exclude irrelevant scenes. For our model, we choose the ResNeXt-50 architecture combined with the adaptive momentum (ADAM) optimization algorithm. We also compare the performance of this ResNet -50 architecture using the momentum-based stochastic gradient descent (SGD) optimizer. Our experimental results show that the ResNeXt-50 model, combined with ADAM, outperforms the ResNet-50 model with SGD in terms of recognition accuracy and robustness. This study provides valuable insights into the application of transfer learning for drone landing scene recognition and offers a more reliable solution for autonomous drone navigation in realworld scenario