Multi-Task Deep Learning Design and Training Tool for Unified Visual Driving Scene Understanding

Abstract

Visual driving scene perception systems have been gained popularity among the autonomous driving research community following the advent of deep learning technology. Moreover, the multi-task deep learning model has been an important tool with respect to unifying the tasks performed in a driving scene perception system, such as scene classification, object detection, segmentation, depth estimation. In this paper, we introduce our developed multi-task deep-learning model design and training tool, for unified road scene perception model. Additionally, we also propose a sequential auxiliary multi-task training method that can train a multi-task model, using different datasets for each tasks. Finally, we present a unified road segmentation and depth estimation model, based on multi-task deep learning, to verify the efficiency and feasibility of our developed tool. Experimental results for KITTI datasets show that our tool-based unified road segmentation and depth estimation model can successfully segment the driving road and estimate its depth.