Efficient Training for Perception Modules in Autonomous Driving

Abstract

Autonomous vehicles aim to drive themselves to their destination without human intervention, and to achieve full autonomy, it is essential to improve perception modules that can accurately acquire information about the vehicle’s surroundings. While many perception modules are showing significant performance improvements due to the development of deep learning technology, there needs to be more research on how to develop efficient modules according to the characteristics of autonomous vehicles. The main characteristics of autonomous vehicles that affect the development of perception modules are limited computing resources and the high cost of building datasets. Therefore, this paper introduces an end-to-end learning method including a domain conversion module for sensor fusion, an efficient perception module learning method using hourglass networks, and odometry estimation methods based on self-supervised learning that can be learned without ground-truth. Each method presents an efficient learning method for perception modules in terms of computation and data and shows that high performance can be maintained simultaneously.