Application of Deep Neural Network Structures in Semantic Segmentation for Road Scene Understanding

Semantic segmentation is crucial for autonomous driving as the pixel-wise classification of the surrounding scene images is the main input in the scene understanding stage. With the development of deep learning technology and the impressive hardware capabilities, semantic segmentation has seen an important improvement towards higher segmentation accuracy. However, an efficient sematic segmentation model is needed for real-time applications such as autonomous driving. In this paper, we discover the potential of employing the design principles of two deep learning models, namely PSPNet and EfficientNet to produce a high accurate and efficient convolutional autoencoder model for semantic segmentation. Also, we benefit from data augmentation for better model training. Our experiment on CamVid dataset produces optimistic results and the comparison with other mainstream semantic segmentation models justifies the used approach.