Autonomous simulated vehicles for steering angle prediction

Abstract

At present, vehicles have many advantages for transporting people to communicate and travel to other places. However, many statistical reports represent over 80% of accidents mostly occur from human error. In addition, the increasing cases of accidents were found from the missing control of the steering wheel. Therefore, this research will develop a new system to reduce human error to decrease the accident case from transporting by suggesting a suitable steering angle. This research proposes a new convolutional neural network (CNN) model for predicting the efficient steering angle based on the imaging view of the vehicle. In the first step, our research gathers the training, validation and testing data from the driving simulation of the Udacity platform that consists of the sequence of driving images, steering angles, speed numbers and others. Then, we convert all gathered data of driving images by reducing the image size to 320x160 pixels and the colour channel from reg-green-blue (RGB) to hue-saturation-value (HSV). In the third step, this research uses the prepared data to train the CNN system for making the prediction model of a suitable steering angle. In the experimental results, this CNN model was evaluated by using the Root Mean Squared Error (RMSE) and R-square value. In addition, we evaluate the percentage accuracy of correcting angles between predicted and gathered steering angles by using the Ground truth degree formula. From the experimental result, this proposed CNN model represents the corrected accuracy of the predicted steering angle of more than 80%. Therefore, this CNN model can use the suggestion of a suitable steering angle based on the real vehicle view for decreasing the human error.