This paper aims to train the driving behavior model through reinforcement learning, which has recently attracted attention as a core technology for autonomous driving. First of all, during the construction of the ANN-based model, the input data was composed of sensor information without visual information, and the reward system was designed to obtain higher reward as the vehicle is centered on the road and the direction of the road coincides with the direction of the vehicle. As a result of learning, the model showed the result of completing the track, so the effect of driving behavior modeling through reinforcement learning was proved. However, the ANN-based model did not include visual information, so it was necessary to select driving behavior based on the current situation.
Therefore, we would like to propose a CNN-based model that determines driving behavior along with driving information as well as status information. First, a future prediction model was designed to predict the future situation rather than the present. This is a model that predicts the future driving image through the past driving image. In addition, a corner classification model was designed to recognize the created future and present situation. The data was constructed by fusing information from these two image processing models with sensor information. In addition, the conditions of reward system are designed to obtain high reward when driving along the in-course in a corner section.
In this paper, in order to compare the performance of the image-based driving behavior model and the existing state-based driving behavior model, the driving distance, driving speed, and driving time were measured for three corner sections on the same track. As a result, it was confirmed that the driving distance was reduced in both sections except for one section, and the driving speed was increased and the driving time was shortened in all sections.