A real-time simulation environment architecture for autonomous vehicle design

Abstract

Various proposed approaches for autonomous driving basically involve an image processing and a machine learning process. It is extremely important to use appropriate image processing techniques and a comprehensive data set in these approaches. Moreover, the proposed model must work in real-time. On the other hand, designing and manufacturing an autonomous vehicle model results in serious hardware costs. In addition, the design and manufacturing processes need to be repeated to develop new approaches. In this context, utilizing a real-time simulation environment can be seen as a suitable approach for a less costly prevalidation of such models. In this study, a real-time simulation architecture is developed with Unity framework to test an autonomous driving model. In addition, an autonomous driving model that includes lane tracking and object recognition approaches is proposed, and an autonomous vehicle simulation is created. Finally, the feasibility of the proposed simulation architecture is tested with the convolutional neural networks-based YOLO algorithm and R-CNN algorithm versions. According to the findings, it is observed that Faster R-CNN, Mask R-CNN and YOLO-v4 algorithms produce results with 91%, 93% and 95% accuracy, respectively. It has been determined that these results are close to the accuracy rates obtained on different traffic sign data sets in the literature. Considering the outcomes, it is argued that a vehicle simulation with an autonomous driving model has been successfully tested in the proposed system architecture.