Environment perception in racing simulators using Deep Neural Networks

In this paper, we introduce a real-time approach for object recognition, applying the Faster Region-based Convolutional Network (Faster-RCNN) concept in an affordable and open source simulator that can be used to train and test Deep Neural Networks (DNN) models. The objective is to provide the scientific community with a framework where the advantages and disadvantages of already published or new designed architectures and concepts for object detection can be validated. The framework proposes two available configurations for implementing DNN algorithms. Our framework provides an interface to the TORCS racing game simulator, where the deployed methods can be evaluated. Various ground truth information can be extracted from TORCS, like the traffic scene image, car position on the track, speed or distances between the traffic participants.