Map Information Updating via Neural Inference for Fast VSLAM

The patent replaces the traditional gradient-based mechanical optimization system (steepest descent and Gauss-Newton methods involving Hessian matrix calculations) with an inference-based system using neural networks. The neural network learns the complex relationship between reprojection errors and map information updates during training, then can quickly predict optimal updates without performing iterative gradient calculations during actual operation, thus substituting complex computational mechanics with a simpler inference process.

The patent applies preliminary action by pre-training neural networks offline using extensive simulation data and labeled examples of reprojection errors and corresponding map information updates. During actual operation, the trained neural network models are deployed to quickly infer optimal map information updates without performing time-consuming iterative calculations, effectively performing the complex learning work in advance before real-time operation.

The patent substitutes the iterative mechanical optimization process (requiring repeated calculations of reprojection errors, gradient descent steps, and convergence checks) with a single forward pass through pre-trained neural networks. The neural network has learned during training how to map reprojection errors to optimal corrections, enabling direct prediction of accurate map information updates without iterative trial-and-error processing.

The patent creates a computational model (neural network) that copies and internalizes the complex optimization logic learned from training data. Instead of reperforming the optimization calculations in real-time, the system uses the trained neural network model to replicate the optimal correction patterns that were discovered during training, significantly reducing processing time while maintaining accuracy.