Indoor Navigation Based on Model Switching in Overlapped Known Regions

Abstract

This paper introduces a novel drone navigation algorithm based on overlapped known regions (OKR). Each OKR has associated a neural network model, which takes as input an RGB image from a camera located at the top of the drone. This model generates two outputs: the distance to the center of the region, and the orientation of the vector that points to the center of the region in the horizontal plane. These regions are constrained to overlap the center of neighbor regions. After training, the drone is able to navigate continuously through several regions by switching the model parameters once the center of each region is reached. Additionally, in order to significantly reduce the number of parameters of each model an adaptive convolutional kernel (ACK) is used, which is able to redefine the convolutional kernel during the inference time according to the input image.