Akıllı kentsel sahne analizi için derin öğrenme mimarileri

Smart city analysis becomes an emerging field in autonomous urban life problems. Image semantics is a complex problem where the image classification, the semantic segmentation and the object detection subroutines are staged in a cascade framework through spatio-temporal datasets. Urban scene analysis has been coupled in several applications such as aviation, robotics, city security, autonomous vehicles, and mass transport. The initial problem of an urban scene is characterized as the pursuit of discrete 2D-3D movements on the streets of a city. Therefore, an accurate segmentation of the scene is required to minimize the spatial gap in the scene. The complete framework provides critical decision-making tools to protect the human beings and the moving objects. To prevent the accidents, new generation autonomous systems turn on their real time sensors to monitor all possible movements in a street. The task of semantic segmentation is to label every pixel including the background into a semantic class. The object detection locates the presence of objects with a bounding box and types or classes of the located objects in an image. Therefore, the detection requires the invariant representations whereas the segmentation needs the equivariant representations. The instance segmentation contains these two tasks: object detection and semantic segmentation. This thesis is composed of the instance segmentation of base objects through Cityscapes dataset. YOLACT deep learning architecture has been applied on high resolution images. The method has been found fast as it requires one stage segmentation. The quality of the masks was better for the large-scale objects.