VISION BASED 3-D LOCALIZATION OF UAVS USING DEEP IMAGE MATCHING

Abstract

Drones and UAVs are used both professionally and privately and they rely heavily on onboard GPS, altimeter, survey meter along with other navigational equipment for completing their automated flight paths. These systems are prone to failure as well as jamming, hence they require a backup plan to complete their flight. We focus on using computer vision based techniques to help UAV determine its height and location to complete its flight.Our objective is to use this technique and find the height as well as location of the UAV and hence 3D localize it. First contribution is gathering/ collection of own custom dataset which is used for the thesis. This dataset was collected by flying Phantom 4 Pro drone in the premises of Nust Main campus. Ten sorties were flown on different heights during day timing to collect data. The dataset consists of several images from ten different heights. The dataset collection height starts from 100 meters all the way up to 280 meters. Flights were 20 meters apart from each other. Second contribution is designing a network which estimate the height of UAV using pair of images. Pyramid Stereo Matching Network is used as base network to extract the features from image pair and then further used to find the disparity map. This disparity map is then further processed and convolved, and height is estimated after refinement. This model uses different Computer vision techniques and convolution layers to estimate the results.