Ensuring Safety of an Autonomous Vehicle with a Neural Network Perception

The continuous technology development both in the field of simulations and neural networks is making Autonomous Driving (AD) more and more appealing for researchers. In order to make the AD a reality that would change our lives on a daily routine basis, the work is divided into two main branches: computer vision and control algorithms improvement. The ultimate goal of the research community is to improve the safety of the autonomous vehicles so that they can overcome the already very high safety of human drivers. The goal of the work is to monitor an autonomous vehicle exploiting a set of Neural Networks for the perception of the surrounding and the prediction of the other traffic participants behavior in order to improve the safety. The scenario is developed using the open source simulator CARLA. The autonomous vehicle is controlled by a PID controller and is equipped with a Radar for the data perception. The Neural Networks are two feedforward networks responsible for the classification of the data and for the prediction of the future radar readings, and a LSTM network responsible for the prediction of the next state of the autonomous car. The behavior of the other traffic participant, which in this case is a leading vehicle, is implemented to be random, and as a consequence the prediction of its next state comes from a random FSM. All the probabilities used for the FSM are arbitrarily assigned trying to reproduce the human behavior in the real world.