Abstract:
With technological advancements in automation and artificial intelligence, automated vehicle
systems, fully autonomous vehicles, self-driving cars are becoming the part of news
constantly. In order to assist drivers in critical moments, for fast response and decision
making, a methodology is required to conciliate the burden on human drivers. A recent
research by MIT shows that number of fatalities can be lessen to as much as 1% by using
Autopilot (self-driving cars). In order to provide full alternative to human driver CNNs are
being used to steer the car perfectly and understanding the internal representations of the
required processing steps on its own without optimization of human-selected features like
detecting the outline of road, lane-following, traffic signals etc. This thesis work targets to
propose an end-to-end solution in which images will be given as an input to CNN model and
the output is steering angle. The datasets being used are provided publicly by autonomous
driving startups. Initially two main designs were evaluated i.e. Nvidia Approach (real-time
environment model) and Udacity Approach (simulated environment model). The major
network proposed in this thesis performed training using real-time dataset and testing was
done on simulated dataset. So it overcomes the problem of state of the art implementations
like over-fitting and domain adaption. During experimentation it is observed that domain
adaption model outperforms with an accuracy of 97% approximately and a loss of 0.017. The
network trained on real-time dataset was able to drive in simulated environment’s
representative track without any inconvenience. However, an even more interesting approach
would be to perform network’s training on simulated (artificial) driving dataset and
testing/validating it on real-time dataset. This would be beneficial since the networks can be
trained on data that can be effortlessly and cost effectively produced in huge quantities.
