SuperPoint and SuperGlue-Based-VINS-Fusion Model

SuperPoint and SuperGlue-Based-VINS-Fusion Model

연구 분야: Verification

학회: International Conference on Intelligent Computing

With the arrival of modern technology, research on safety inspection techniques in complex environments is crucial. Using hardware devices such as mobile devices and vision sensors combined with many algorithmic techniques can effectively improve inspection efficiency, reduce costs and ensure safety. Among them, vision SLAM technology and target detection technology play a key role. Vision SLAM technology provides precise positioning through vision sensors, enabling mobile devices to navigate autonomously in complex environments, while target detection technology utilizes deep learning algorithms to accurately identify potential safety hazards for early detection and treatment. In this context, the research work in this paper focuses on improving the localization accuracy of the vision SLAM system in complex environments under complex conditions such as light changes, dynamic scenes and weak textures, so as to enhance the detection efficiency in complex environments. Aiming at the problem of degradation of the quality of feature points extracted by traditional feature point extraction methods in weak texture scenes, this paper proposes a technical solution using SuperPoint for feature extraction and SuperGlue for front-end feature matching. And it is integrated into the VINS-Fusion system. By combining the advantages of SuperPoint and SuperGlue, the VINS Fusion system not only improves the visual localization accuracy in complex environments, but also enhances the system’s adaptability and robustness to dynamic changes, ensuring that accurate and reliable localization can be achieved even when the visual features are not obvious. Achieve accurate and reliable localization even when visual features are not obvious. In this study, the fusion SuperPoint feature extraction and SuperGlue feature matching in VINS-Fusion are successfully accomplished. The experimental results show that the loopback error of the VINS Fusion system is significantly reduced in the y-axis direction from 0.40 m to 0.17 m, which proves the reliability and stability of the system in real navigation tasks.