Visual SLAM

Visual SLAM estimates a camera’s motion while constructing a map of its surroundings, allowing robots and other systems to navigate without relying entirely on external positioning. Classical systems match image features and refine poses and maps with geometric optimization and loop closure; newer approaches learn visual correspondences or use dense scene representations such as neural fields and 3D Gaussian Splatting. Current work also tackles moving objects, changing illumination, uncertain depth and scale, and specialized settings such as underwater, panoramic, and large-scale outdoor scenes. A central tension is achieving accurate, globally consistent maps in difficult environments while keeping computation and memory low enough for real-time use, with benchmark results not always reflecting deployment performance.