Behavior selector for autonomous vehicles using neural networks

In this work, we propose a method to build a behavior selector for the autonomous vehicle AutoMiny. With this, the vehicle was able to: drive lane-keeping and under the speed limit, pass parked cars, stop if a pedestrian (or another obstacle) appears in front of it, and park when the "passenger" sends a signal. The behavior selector designed was a feed-forward neural network where the inputs are seven binary variables whose values change depending on the sensors' data, and four neurons in the output correspond to four driving maneuvers. To detect and localize traffic signs, other vehicles, and commons obstacles on the road (like pedestrians, dogs, cats, etc.) a convolutional neural network with YOLOv3 architecture was designed, trained, and implemented. With this, it was possible to process the camera's images and define several states for the decision-making neural network. All proofs were realized using the Gazebo-ROS software in a simulator designed for the AutoMiny vehicle.