Vision-based landing site detection for a UAV: From theory to application

This study addresses the critical need to enhance the safety of multirotor UAVs during emergency flights by addressing the challenges of identifying suitable landing spots and avoiding collisions with obstacles. Drones encounter difficulties due to their elevated height, leading to potential collisions, inaccurate distance estimation, and weak radio signals. Researchers are diligently developing a cost-effective and efficient landing system to mitigate these issues and enhance overall safety. Key gaps in vision-based landing site detection for UAVs include adapting to varying environmental conditions, achieving real-time processing, precise obstacle recognition, ensuring robustness, and scalability, handling data anomalies, integrating with other sensors, obtaining diverse training data, complying with regulations, and maintaining cost-effectiveness. To address these challenges, the proposed model combines a U-Net, a deep Convolutional Neural Network (CNN) architecture, with a ResNet 34 backbone. The model builds a labelled aerial photo database using image segmentation techniques for CNN training, determines landing areas through 2D dataset image analysis, and plans paths from the UAV's position to the designated area centre using the A* algorithm. The entire research is implemented in Python, harnessing TensorFlow's capabilities. Impressively, the model achieves a training accuracy of 96.6% and a validation accuracy of 96.34% while utilizing approximately 80% of the dataset, comprising 1480 images, during the training phase, with a recorded training loss of 0.1034. These outcomes underscore the model's exceptional accuracy, establishing it as a cornerstone for subsequent landing area selection and path planning by the algorithm.