Research on Stereo Vision Localization Method Based on Prior Semantic Map

In fields such as intelligent robotics and autonomous driving, visual localization is a fundamental and crucial problem. During long-term operation, the absence of loop closure or prior constraints inevitably leads to accumulated errors. To address this issue, this paper proposes a novel visual localization method that integrates prior semantic maps. The method obtains local depth maps using stereo disparity and performs semantic segmentation on RGB images. The segmented images are aligned with the depth maps, and semantic information is mapped onto the depth maps to generate 3D semantic point clouds. After visual odometry initialization, the method employs the Iterative Closest Point (ICP) algorithm to align visual point clouds with prior map point clouds using invariant semantic information, resulting in more accurate camera poses. This method has the following advantages: By selecting only fixed and invariant objects for registration, the system demonstrates strong robustness against environmental changes while relying solely on low-cost, lightweight camera sensors to complete the tracking task, significantly reducing the overall system cost. Experiments conducted in real-world scenarios validate that this method can estimate camera pose more accurately than visual odometry.