부산대학교 도서관

Simultaneous localization and mapping (SLAM) is computational technique forrobotic system with which it moves in fixed or predefined map having unknownenvironment. The SLAM has an objective toward localization for the robot inunknown environment of given or predefined map. Many learned people havedefined their own way of defining and designing SLAM algorithm in whichmSLAM, vSLAM, FullSLAM and Extended Kalman filter (EKF)-based Gmapping SLAMare prominent. Out of which, we found that EKF-based SLAM algorithm performsbetter. It can be scaled for better précising by adds-on with other curvealgorithm. A mixture or serial operation of these algorithms may lead tobetter optimization of the system. After doing experiments, we stuck withoptimizing or creating better precision for localizing in the environment byapplying few parametric curve algorithm over EKF. To make it more optimized,now we are experimenting EKF with machine learning (ML)-based optimizationfor SLAM. SLAM algorithm enables computer systems predict and update robotposes when robot is moving into given map and helps in localization. Butthese were not very precise in nature due to mathematical approximation ofprediction, to overcome it a ML approach may be applied for betterprecision. In this chapter, we will discuss artificial neural network(ANN)-, k-nearest neighbor (kNN)-, CNN-based optimization techniques foroptimizing precious in localization applied on EKF-based SLAM algorithm. Toimprove precision, an argumentation of ML can be examined, in this respect,we are designing and testing a ML system based on EKF. The systemincorporates with argumentation and learning modules based on ML and deeplearning. To evaluate the effectiveness of the proposed learning toprediction model, we have developed the ANN-based learning module. For themodel trained with deep ML modules, the machine readjusts itself duringnavigation of robot on different surfaces (marble flooring, boulder tile,rough tile) due to which friction gets controlled and get better pathprediction results. Various optimization algorithms are available in deeplearning among which kNN, ANN is prominent one which is in great use. Out ofwhich ANN is shallow network which is able to track complex SLAM problem andlatter one has lower accuracy for tracked mobile robots. The deep learningsystem based approach shows significant increase in performance.