LRODNet: a framework for 3D object detection of railway tracks in low-light environments
TL;DR
An attention based multi-stage fusion module, which achieves adaptive feature alignment and deep fusion through attention mechanisms is proposed, providing an efficient and reliable solution for environmental perception of railway tracks in complex lighting environments.
Abstract
Accurate detection of railway tracks at various scales in low-light conditions is a critical challenge for ensuring the autonomous and safe operation of rail trains. To address this, this paper proposes LRODNet, a framework for 3D object detection of railway tracks under low-light environments. Specifically, we design an image feature extraction module that efficiently extracts image features through a lightweight convolutional block and a feature adaptive enhancement module. To tackle issues such as low contrast and significant noise in low-light images, the feature adaptive enhancement module is developed to effectively enhance detail representation in key regions while suppressing background interference. Addressing the difficulty of cross-modal matching between image features and LiDAR features, we propose an attention based multi-stage fusion module, which achieves adaptive feature alignment and deep fusion through attention mechanisms. Experiments demonstrate that LRODNet exhibits outstanding performance in detecting railway tracks under low-light conditions, providing an efficient and reliable solution for environmental perception of railway tracks in complex lighting environments.