A Pioneering Scalable Self-driving Car Simulation Platform

This paper proposes a novel hybrid, cross-platform 3D self-driving simulator, which offers a platform for researchers in the field of automatic vehicles to validate algorithms with ease and criteria to evaluate models in the community. Primarily our system decomposes the achievement of self-driving into four distinct steps by using a Tensorflow framework in machine learning: data generating, data balancing, model training, and model testing. The core part of our training section, the Alexnet, underlies training a deep neuron network with high accuracy at a fast speed. Some techniques, like balancing data, are employed to avoid overfilling, thus improving the robustness and accuracy; other functions, such as speed control, are designed for emulating the reality more vividly. By many study cases, it shows a strong flexibility of our simulator substantially. Various lane detection and obstacle detection algorithms are embedded smoothly in our system. In addition, Canny algorithm and Hough Transfer algorithm detect and draw out the edge of lanes perfectly, while YOLO architecture manages to detect obstacles.