Multitask learning for autonomous driving

Abstract

Autonomous driving is inherently a multitask learning (MTL) problem. In the current work, we propose a generalized MTL framework for the estimation of various parameters needed for autonomous driving. This framework generates different networks for the estimation of a different set of tasks based on their relationship. The relationship among tasks to be learned is handled by including shared layers in the architecture. Later, the network separates into different branches to handle the difference in the behavior of each task. More specifically, we provide a solution for the estimation of driving control parameters as well as those related to scene information. We demonstrated the performance of the proposed solution on four publicly available benchmark datasets: Comma.ai, Udacity, Berkeley Deep Drive (BDD) and Sully Chen. A synthetic dataset GTA-V for autonomous driving research has also been proposed to further evaluate the proposed approach.