Construction of Topological Navigation Map Based on Model Fusion

With the advancements in artificial intelligence technology and the field of mobile robotics, small robots like drones and unmanned cars have become increasingly visible. However, the trajectory maps created by unmanned vehicles in complex environments pose challenges for planning and navigation tasks, often leading to significant misidentification issues during the mapping process. To tackle these problems, we propose a method for constructing topological navigation maps based on model fusion. This approach aims to reduce the misidentification rate through model fusion, thereby enhancing the efficiency of constructing topological maps. In this study, we utilized the ORB-SLAM2 framework to extract keyframe information from a dataset of unmanned driving videos. These keyframes were fed into a deep learning neural network for intersection recognition. To improve recognition accuracy, we employed model fusion by averaging multiple points on the SGD trajectory using SWA. As a result, we were able to generate a topological map. Our experiments confirmed the effectiveness of the model fusion-based intersection recognition network and the construction of the topological navigation map. Furthermore, we evaluated the fusion method on common benchmark datasets and found that our results were competitive compared to the current state-of-the-art approaches.