Image segmentation architectures and their practical application for unmanned aircraft landing

This paper discusses image segmentation architectures and their practical application for unmanned landing, examples of implemented systems for delivering goods by drones. The optimal drone landing was selected and implemented using image segmentation by a neural network. The data set for training is selected and modified by mirroring the image and rotating the image. An additional data sample was created using an application that simulates the flight of a drone in a custom environment. Found a way to quickly mark up a video recording of a drone landing. Neural networks with different architectures were tested to find the best one for needed purpose. For this task four trainings of different models were done to find needed accuracy and speed of analyzing one frame of a video. Two different metrics of accuracy were used for neural network to get reliable data of recognition. The best environment classes of recognition were selected. Three models with different sets of classes were created. Results were analyzed and the best classes were selected. The final model of the neural network was done on the basis of linknet architecture. For best performance additional pretrained weights were chosen. Training of neural network had been going for 13 hours Video of a drone landing was processed in a virtual environment. Based on height of a drone landing spot and area of following segmentation was chosen. The optimal height above the surface to start the landing process. Was made a demonstrative video of drone landing in optimal conditions.