Multi Camera System Analysis for AutonomousNavigation using End-to-End Deep Learning

This work aims to present an autonomous vehicle navigation system, based on an End-to-End Deep Learning approach, and to study the impact of different image input configurations to the system performance. The proposed methodology in this work was to adoptand test different configurations of RGB and Depth images captured from a Kinect device. We adopted a multi-camera system, composed by 3 cameras, with different RGB and/or Depth input configurations. Two main systems were developed in order to study and validade de different input configurations: the first one based on a realistic simulator and the second one based on a mini-car (small scale vehicle). Starting with the simulations, it was possible to choose the best camera/input configuration, then we validated that using the real vehicle (mini-car) with real sensors/cameras. The experimental results demonstrated that a multi-camera solution, based on 3 cameras, allow us to obtain better autonomous navigation control results in a End-to-End Deep Learning based approch, with a very small final error when using the proposed camera configurations.