Enhancing a Trinocular Depth Camera Navigation Environment and Expanding a Realistic Scene Dataset with Horizontal Lines

This paper presents the creation of a real-world trinocular stereo vision dataset featuring horizon-line obstacles, captured using the Intel RealSense Lidar Camera L515 for depth ground truth. Current open-source real-world stereo vision datasets do not provide a trinocular camera setup. Therefore, we developed a dataset that includes real-world horizon-line obstacles to enhance our model's ability to recognize horizontal lines in practical scenarios. Additionally, we constructed scenes that simulate potential navigation challenges, such as close-range and complex environments, enabling our model to be applicable in future drone navigation tasks. Furthermore, we expanded ground truth coverage by a dilation method.