Visual Inertial Odometry with Fusion of Point and Line Features in Low Illumination Environments

摘要

The increasing sophistication of robotic intelligence has spurred a growing demand for autonomous navigation and localization technologies, particularly in dimly lit and complex environments. Visual inertial odometry (VIO) has emerged as a popular solution owing to its simplicity and cost-effectiveness. However, traditional point feature tracking methods often suffer from diminished accuracy in challenging conditions such as low-texture environments and variable lighting, posing significant obstacles to effective positioning and navigation for intelligent devices. In this paper, we introduce rich line features into dim environments based on existing point feature tracking, improve feature extraction and matching algorithms, and construct a graph optimization nonlinear model that integrates point-line-IMU information fusion. To mitigate excessive line segment segmentation, we introduce a gradient density filtering mechanism to extract prominent line features, followed by a process of fitting and merging similar broken lines based on angle characteristics and spatial relationships between segments. Leveraging factor graph models, we formulate a comprehensive graph optimization framework for multi-state pose estimation, enabling a positioning system in dim environments that amalgamates point and line features with inertial navigation. Comparative analysis with conventional VIO methods relying solely on visual point features and inertial sensors underscores the superior accuracy and robustness of our proposed approach. Notably, during a 120-meter trajectory, our algorithm achieved a reduction of approximately 3 meters in the maximum positioning error.