Development of perception system for autonomous vehicles by using GPU-based learning

Autonomous vehicles are that perceive environmental conditions and act in accordance with decisions. At present, interest in autonomous vehicles is increasing rapidly. With the development of sensor and GPU (Graphics Process Unit) technology, and innovations in artificial learning methods, autonomous vehicle technology is also developing. In this study, firstly, general information about autonomous vehicles was given. Secondly, using the Caffe in the Linux environment, the MNIST dataset was used for digit recognition. For this application, two computers were used in addition to the Nvidia TK1 and Nvidia TX1 cards, and the CPU (Central Process Unit) and GPUs were evaluated for speed and accuracy. Thirdly, object recognition applications have been implemented. For this purpose, firstly, in Matlab in Windows environment, pedestrian crossings traffic sign detection application was done. The originally created dataset was used for this application. For detection, Region Convolution Neural Network (R-CNN), Faster R-CNN was used and the results were examined. Later in the Linux environment, object recognition was implemented using Yolo in a basic sense and output was given. Finally, a small autonomous land vehicle was built using artificial learning methods. For this purpose, various sensors, camera and Nvidia TX2 card were installed on the land vehicle. The autonomous vehicle were tested on the designed racecourse. All applications were realized successfully. The results were given comparatively in tables and figures.