Camera-Based Online Vectorized HD Map Construction With Incomplete Observation

Abstract

Camera-based online map construction focuses on learning map elements from surround-view images. Distinguished with previous methods that rely on complete observations, we explore a new map construction problem under incomplete observations where one or more perspectives of the surround-view are missing due to camera damage or occlusion. Incomplete observations lead to inferior performance and may even result in failure. Map construction based on incomplete observations faces two challenges: supplementing missing perspective features and reducing the complexity of high-dimensional feature learning. To address these issues, we propose a novel <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Panoramic Observation Prior Network</i> (POP-Net). Firstly, based on the observation switch training mechanism, we propose a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Panoramic Learning Module</i> (PL-Module). It establishes a learnable panoramic feature space, facilitating the extraction of panoramic features from incomplete observations, thus supplementing missing perspective features. Secondly, based on the feature decomposition mechanism, we design a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Panoramic Decomposition-Aggregation Operation</i> (PDA-Operation), which decomposes high-dimensional panoramic features into low-dimensional local scene features. This allows limited local scene features to represent diverse panoramic features, alleviating computational and memory burdens of high-dimensional feature learning. Experimental results demonstrate that our method surpasses existing approaches under incomplete observation scenarios.