Mobile Robot Localization With Local Maps and Smooth Map Switching

The patent divides the global map point clouds into multiple local maps, each covering a specific region. Only the current local map is loaded into memory for localization operations, while other regions remain stored on disk. This segmentation approach maintains localization completeness for the current region while dramatically reducing memory consumption compared to loading all point clouds simultaneously.

The patent pre-configures transition regions between adjacent local maps during map construction. These transition regions contain overlapping point cloud data from neighboring areas. When the robot approaches a map boundary, the system proactively loads the next local map into memory using the current pose, ensuring a smooth transition without localization jitter. This preliminary preparation of transition data eliminates the instability that would otherwise occur during map switching.

The patent extracts and retains only key frames from the complete point cloud data, discarding redundant intermediate frames. These key frames are strategically selected to represent important spatial locations and features. During pose matching, only these extracted key frames are used for comparison, maintaining localization accuracy while significantly reducing the computational burden of processing all original point cloud data.

The patent implements a dynamic multi-threaded computation architecture where the system automatically adjusts the number of threads based on operational conditions. During normal localization, multiple threads perform parallel operations such as point cloud registration, pose estimation, and map updates simultaneously. The system dynamically manages thread creation and termination based on computational load, maintaining high operational efficiency while adapting to varying system states and resource availability.