Navigation Aids Based on Optical Flow and Convolutional Neural Network

Unmanned Aerial Vehicles (UAV) have been employed in various activities, such as search and rescue missions. Global Navigation Satellite System (GNSS) is the main tool used by UAVs to identify their global location and to be able to complete flights safely. However, UAVs can suffer attacks that invalidate the global positioning information or simply loose for a period the GNSS signal. The lack of the aircraft's global positioning information can result in incomplete missions and accidents. In this paper we propose an Image-Based Localization System (IBLS), which allows the global position of the aircraft to be inferred based on images captured by a camera pointed at the ground. Our proposal is based on the concepts of optical flow to infer the displacement between two images captured sequentially with a CNN. Based on the displacement, we use the haversine formula to estimate the new global position (latitude and longitude) of the UAV. IBLS uses a calibration step learned during flight with available GNSS signal, which allows to reduce the error in inferring the new position. The results achieved on simulated and real data sets demonstrate that our proposal is able to infer the position of the UAV with lower error than in GNSS. The main contribution of this work is the investigation of a system based on the concept of Optical Flow and Convolutional Neural Network to estimate the geographical coordinate during UAV flight.