With the development of machine vision technology, there have been new developments in the study of the subject of autonomous driving. Differences in autonomous driving capabilities have become one of the core competencies of car manufacturers today. Nowadays, laboratories and universities worldwide are constantly trying to break through the influence of the real environment on autonomous driving and improve the accuracy of predictions and the rationality of decisions on autonomous driving. In this context, and taking into account the current state of development of autonomous driving technology, this paper will mainly explore the prediction and decision-making of autonomous driving based on deep learning throughout the process of autonomous driving, mainly from the camera as the input side in the first stage, and later on, we will add some other sensors to allow better fusion of multiple sensors through optimized models. We will use DonkeyCar, an open-source autonomous driving platform, as our basic autonomous driving development platform, to optimize and improve the accuracy of prediction and rationalization of decision-making from hardware to software level by combining deep learning. This will enable it to have a deeper level of environmental awareness throughout the autonomous driving process.
In this section, we focus on the discussion about deep learning-based autonomous driving simulation vehicles. We will start with the experimental design related to our research approach, which includes aspects of autonomous driving system integration, such as hardware and software, as well as data acquisition and data pre-processing, training the collected data, and evaluating the trained models. Our research in this case is based on the small open-source autonomous driving platform Donkeycar. donkey car includes a comprehensive set of software tools for training and testing autonomous driving models. It uses the popular TensorFlow and Keras libraries for deep learning, allowing developers to build and test their neural network models. The platform also includes tools for collecting and pre-processing data, as well as visualizing the performance of trained models.
