Innovative solution approaches based on deep learning for autonomous driving

Advances in deep learning have led to significant developments in autonomous driving technologies in recent years. These developments have aroused great excitement in various industrial and academic fields and accelerated the development of autonomous vehicle technologies. The use of deep learning methods has played a critical role in improving the basic capabilities of autonomous vehicles such as environmental sensing, decision making and lane following. Many autonomous vehicles developed today utilise a rich sensor infrastructure, often equipped with costly sensors such as LiDAR, radar and advanced cameras. However, without the need for expensive sensors, human drivers can achieve impressive driving performance in daytime, nighttime and even in extreme weather conditions using only their visual perception. Both autonomous vehicles and human drivers generate motion control commands based on their environmental perceptions. Among these commands, steering angle and speed estimation are the main tasks of autonomous driving. This dissertation presents an end-to-end method for estimating steering angle and vehicle speed using various meaningful cues extracted from a monocular camera image. Three input images are used in the proposed model. In addition to the colour image, which conveys scene texture and appearance details, the monocular depth image and the semantic segmentation image, which provide information about spatial and semantic structures of the environment, respectively, are also included. In addition, LSTM units are used to obtain temporal features. The proposed multi-modal multi-task model is evaluated and compared with existing methods on the Udacity and Sully Chen datasets generated by human drivers in real-world conditions and on the dataset collected through the CARLA Driving Simulator. As a result of these comparisons, it is observed that the proposed model achieves the state-of-the-art results in the literature in RMSE values by providing 44,96% improvement for steering angle and 4,39% improvement for speed in the Udacity dataset. Similarly, the results obtained on the Sully Chen and CARLA datasets are the most successful results in the literature. A comprehensive ablation study was also performed to evaluate the effectiveness of each component of the model and to examine its contribution to the model. The results obtained emphasise the potential of autonomous driving systems using only visual input. At the end of the research, a comprehensive dataset has been created for researchers aiming to work in the field of autonomous driving for various purposes. The obtained dataset has become the most comprehensive CARLA dataset in the current literature in terms of sensor diversity, collected data quantity, and data variety. This dataset has been made available ready-to-use, allowing researchers to utilize it immediately upon accessing the data. This approach enables researchers to work quickly on the CARLA dataset without spending significant time and effort to create a dataset.