Developing deep learning-based methods for visual perception of road and traffic elements in autonomous vehicles

A significant portion of traffic accidents worldwide are caused by human errors. Autonomous driving systems are being developed for purposes such as reducing or preventing traffic accidents, reducing emissions, transporting disabled people, providing affordable transportation services to passengers, and reducing driving-related stress. Accumulated knowledge in vehicle dynamics, breakthroughs in computer vision caused by the emergence of deep learning, and the availability of various sensor technologies are catalyzing research on autonomous driving systems. However, to achieve safe driving in urban environments with complex traffic, modules such as localization, perception, planning and vehicle control in the software architecture of the autonomous driving system must be designed perfectly. In particular, the perception module becomes even more important as it is the first phase in which the vehicle interacts with its environment. The result of an incorrect object recognition and tracking in this module directly affects other modules and may lead to the autonomous vehicle making wrong decisions and thus causing accidents. Therefore, in this thesis study, we focus on increasing the accuracy performance of the perception module's tasks such as crosswalk detection, detection of cracks on the road surface, detection of objects in the traffic scene and detection of drivable road regions. For each task, a new method based on convolutional neural networks has been developed and proposed. A new and original data set has been prepared to eliminate the lack of data sets in the literature for the detection of drivable road regions. In experimental tests, when the proposed methods and state-of-the-art methods for each task were evaluated individually, the proposed methods were more successful in terms of accuracy than the state-of-the-art methods.