Alçak irtifada uçan İHA videolarında örtüşme altında görsel nesne takibi

Aerial imagery from flying unmanned air vehicles has been very important for many applications such as satellite remote sensing, object detection and tracking. Visual problems in videos from unmanned air vehicles have higher complexity compared to videos taken from ground-based sources.Moreover, if the videos are taken from low-altitude, the visual problems are more severe; thus, processing of such videos is more challenging. This thesis concentrates on this challenging problem and focuses on target tracking under occlusion, which occurs when a target becomes invisible due to blocking by other moving or static objects. First, a new method, named as PFCNN is developed to track vehicles against most visual problems, except occlusion. PFCNN combines a particle filer with a convolutional neural network; and this combination creates a powerful method for tracking against most of the visual problems. Computer simulations verify that PFCNN has superior performance compared to deep learning-based state-of-the-art methods. Then, PFCNNOcc and final PFCNNOcc,two novel methods, are developed to achieve robust target tracking, especially against short- and/or long-term partial and full occlusion. PFCNNOcc architectures also combine convolutional neural network models with particle filters in order to predict optical flow, occlusion maps, segmentation masks, and classification scores, as well as the degree of occlusion. They use the following functions which are proposed in this thesis: 1. Re-Identification, 2. Occlusion Detector and Calculator, 3. Full Occlusion Dense Search, 4. Merged-Target Detection. Computer simulations verify that these functions lead the final PFCNNOcc to the top of the state-of-the-art Siamese network-based methods. A novel synthetic dataset, named S-UAV, is generated to train and to test the proposed architectures, and it fills the gap, arisen for such a dataset in the literature.