Abstract:
Autonomous systems, including autonomous cars, have become a prominent research
topic in recent years due to their potentials for improving safety for users and the potential
commercial success for automotive industries. The increase in studying autonomous vehicles is
recently motivated by the advancements in machine learning. One of the most challenging
tasks of autonomous vehicles is self-navigation, which dynamically makes decisions for
movement based on the learned surrounding environment. Although many efforts have been
denoted to improving the self-navigation, most of the efforts focus on solely using deep neural
networks (DNN) to analyze the environmental images captured by the video sensors on the
vehicles. This, along with best practices of gathering data and labelling the data for training,
have led to high prediction accuracy. There, however, exists many undiscovered ways to
optimize performance of nontrivial tasks that must navigate dynamic environments. In this
paper, a hybrid deep learning network model composed of both convolutional layers and Long
Short-Term Memory (LSTM) has been proposed to learn an environment in images by
exploiting the time-series imaging data and accordingly make navigation predictions. This leads
to a significant improvement in accuracy in comparison to models that do not exploit temporal
information in the environmental imaging data.
