Use of convolutional neural networks for autonomous driving maneuver

In this work, we use a convolutional neural network (CNN) to process the lidar data of an autonomous vehicle and so get the steering angle to carry out the obstacle evasion and parking maneuvers. To introduce the lidar data and other measurements in a CNN, we map the 400 polar vectors (rho(i), phi(i)) in a 20 x 20 normalized matrix; which the position of each element correspond to an angle phi(i) and the elements are rho(i)/rho(max). We probe the method in simulator developed by the Freie Universitat Berlin [1], getting a similar performance as a finite state machine, used as an expert driver in the training.