Dataset Generation Process for Enhancing Depth Estimation Network in Autonomous Driving

To ensure the safety of autonomous vehicles, accurately perceiving the spatial information of the surrounding environment is crucial. Supervised learning-based camera depth estimation networks can be used for this purpose. However, training these networks requires high-quality depth datasets, but existing datasets have quality problems. This paper introduces a novel dataset generation process that uses Light Detection and Ranging (LiDAR) mapping to enhance the quality of depth datasets. The method consists of three main stages: LiDAR Point Cloud Accumulation, Static Background Depth Rendering, and Dynamic Object Depth Rendering. First, multiple LiDAR scans are collected to build a detailed map of the environment, gathering more comprehensive spatial information than a single scan can provide. Next, depth images of the static parts of the environment, such as buildings and roads, are created to ensure accurate representation of these elements. Finally, moving objects, such as cars and pedestrians, are identified and handled separately to reduce noise and improve the clarity of the depth images. Experiments with a simulation dataset and two depth estimation networks showed significant performance improvements across all evaluation metrics when trained with our proposed dataset compared to existing datasets. Specifically, both networks demonstrated over 60% improvement in SILog and over 40% improvement in Abs Rel metrics. These results demonstrate the effectiveness of the proposed dataset generation process in providing superior training data, thereby enhancing the accuracy and reliability of depth estimation in autonomous driving applications.