Development of lane following and control system in autonomous vehicles using machine learning

Today, the automotive industry is in a rapid development towards autonomous vehicle technology. Autonomous vehicle technology is basically designed to prepare a safe driving environment and to reduce traffic accident rates caused by driver errors. Studies are carried out for this purpose in the field of autonomous driving technologies. Within the scope of this purpose, lane detection and tracking is one of the important areas in autonomous driving technologies. When we look at the studies on this field, it is seen that mainly image processing techniques are used. However, two main problems are encountered while applying image processing techniques for lane detection and tracking. First of all, it is necessary to work with a certain area on the image in order to reduce the processing load in the image and to work towards the right area. Region of interest (ROI) processing is often used to filter the area to be studied from the image. However, since the region is determined with fixed coordinates for this process, it causes restrictions on the working area in cases where the vehicle has to turn. With the Mask R-CNN algorithm used in the study, the ROI region is constantly updated with the new coordinates obtained from the segmentation of the path for each frame. The second problem is that the weather conditions are very effective in the detection of stripes using image processing techniques. Serious problems arise in image processing and detection processes from cloudy, sunny or instantaneous changes in the weather. In the study, a solution to the problem of lane detection in different weather conditions was proposed with the originally developed data set. In addition, with the developed method, deep learning methods were used as a solution to these two basic problems. By using Mask R-CNN and Faster R-CNN algorithms together, these two basic problems have been eliminated and successfully applied for lane detection and tracking. The problem, which is solved by two algorithms, has been experimentally tested on an improved autonomous vehicle. In the model training carried out for experimental tests, both the originally developed data set and the KITTI data set, which is frequently used in the literature, were used separately. The test results prove that the use of both algorithms together in lane detection and tracking for autonomous vehicles is quite successful.