Mapping and artificial intelligence based object recognition with unmanned aerial vehicles

With the developing technology, the number of studies with Unmanned Aerial Vehicles (UAV) is increasing day by day. For example, fire detection, detection of damage after natural disasters and rapid access to victims, in agricultural lands; UAVs are frequently used in many areas such as tracking of crops, spraying, disease detection and monitoring of soil moisture. As in the examples mentioned and in various applications, sensors such as GPS, IMU, LiDAR, camera are vital in extracting the targeted information and various sensors are used. In this dissertation; It is aimed to extract as much information as possible from the sequential images obtained using camera and GPS sensors. For this reason, an approach that includes three different algorithms that can be used for different purposes is presented. In the approach, it is proposed to use deep neural networks, geolocation and image mosaicing methods together to obtain various information in different environments. The approach first started with the deep learning model. In the application, a model named Faster R-CNN (Faster Region-based Convolutional Neural Network) was used and its efficiency in object detection was calculated. After the pixel coordinates of the detected objects were determined, the GPS locations of the objects were determined by the geolocation method. While making this calculation, it is assumed that the camera is constantly demonstrating towards the ground and the height of the UAV does not change. Finally, the images obtained in a certain order were turned into a single and large image with the image mosaic method. It was used in the last stage of the proposed approach so that the errors that may be caused by the mosaicking method do not affect the efficiency of geolocation and object detection. In order for the other two methods not to affect the efficiency of the mosaicing algorithm, the images that are the output of the deep learning and geolocation methods were combined after the necessary geometric transformations were made. Thus, each method has been prevented from affecting each other negatively. In conclusion, in this dissertation study, the detection of objects in the images, the determination of the geographical coordinates of these objects and the merging of sequential images were made for use in different areas.