RGB-D Convolutional Recurrent Neural Network to Control Simulated Self-driving Car

For autonomous vehicles, the environment perception is of great importance given the need to control the vehicle actions in the face of the physical circumstances that surround it. Having more sensors on board increases perception, however the information obtained from them must be of better quality. Therefore, this paper presents an improved Recurrent Convolutional Neural Network to be trained with RGB-D images with depth generated from the fusion of a camera and a LiDAR sensor in a simulated environment in ROS, focused on the steering control of a self-driving car. As a contribution, the simulated environment and an RGB-D database are presented. The experiments demonstrated how the improvement of the neural network achieves better autonomy results by \(9.95\%\) with respect to its base form and the autonomous driving model most cited in the literature. The proposed model offers an autonomy of \(95.9\%\) .