Performance Enhancement using Data Augmentation of Depth Estimation for Autonomous Driving

Abstract

For autonomous driving, various sensors such as cameras, LiDAR, and radar are required to accurately perceive the surrounding environment. These sensors provide information for tasks like object recognition, lane detection, path planning, and distance estimation. However, processing information from these multiple sensors for perception tasks demands significant costs, computational resources, and latency. These challenges hinder the practical implementation of real-time edge computing in autonomous driving systems. Consequently, research is actively exploring methods to perform perception using only cameras, particularly to alleviate the computational burden and cost associated with 3D point cloud data from LiDAR or radar sensors. In this study, we investigate techniques to optimize the performance of Monocular Depth Estimation (MDE) methods, which utilize a single camera to extract 3D information about the surrounding environment. We focus on enhancing accuracy through classical data augmentation techniques and synthetic data generation methods. Additionally, we explore the selection of an optimal loss function. Experimental results demonstrate that employing our proposed data augmentation approach reduces REL by approximately 3.9%, showcasing the potential of this method.