PROJECT TITLE :

An End-to-End Multi-Task Learning Model for Drivable Road Detection via Edge Refinement and Geometric Deformation

ABSTRACT:

In this paper, an end-to-end neural network model is used to present a road detection method that can be used for autonomous driving. Our approach makes use of the qualities of road boundaries as well as the multi-task learning capabilities of a deep convolutional network. In order to improve its performance, we reassigned the label and rebalanced the loss of road pixels. This allowed us to concentrate on the learning of challenging examples on the boundary. After that, a method for data augmentation that is based on road geometric transformations is proposed in order to make the network model more robust when applied to traffic scenarios. On the basis of these two innovative methods, an integrated unified architecture is created. This architecture consists of a shared deep residual encoder network as well as multi-branch decoder sub-networks. It does this by using road scene classification as a supervised learning task so that it can accomplish both road segmentation and scene classification at the same time. The results of the experiments show that the proposed method is capable of achieving the highest possible MaxF value in the majority of the road scenes. Our superior performance on the KITTI-Road benchmark is demonstrated by both qualitative and quantitative evaluations.