Designing Adaptive Filtering For Feature-based SLAM Algorithm Of Wheeled Mobile Robot

Simultaneous Localization and Mapping(SLAM)is a relatively widespread problem that needs to be solved to make a robot fully autonomous.Given the noisy measurement and process,the system architecture should find the accurate position of the robot and construct the map concurrently.Determining the inaccurate position of the robot can makes an improper construction of the map and vice versa.Linguistically,the main objective of addressing the SLAM problem lies in estimating the mean of the robot pose and feature-based map and their covariances.According to this brief description,it is not surprising that the use of the Extended Kalman Filter(EKF)and Smooth Variable Structure Filter(SVSF)have been significantly solving the problem of Simultaneous Localization and Mapping(SLAM).The implementation requires an accurate system model and known prior knowledge.However,the theoretical perspective illustrates no precise system model due to some considerations,such as avoidance of physic laws.Besides that,the prior knowledge of noise statistics is usually unknown or partially known in the real application.Therefore,by manually defining them as the conventional and common way,both the performance of the SVSF and EKF possess a high risk of degradation.The inaccuracy of the modeling system might enlarge the estimation error.The uncertainty caused by the unpredictable and random error might affect the characteristic statistical change,which undoubtedly leads to the filter divergence condition.Hence,the traditional form of both filtering strategies should be initially enhanced.The significant contribution of this research is to equip EKF and SVSF with an ability to estimate the noise statistic of the process and measurement and its corresponding covariances.This strategy is well-known as an adaptive filtering method based on a batch estimation of parameters.It is a popular method to tune the gain by offlinecalculating the unknown parameter.Henceforth,they are termed as Adaptive Extended Kalman Filter and Adaptive Smooth Variable Structure Filter(AEKF and ASVSF).As an effort to accomplish these goals,in the first case,the conventional EKF and SVSF are respectively derived by referring to the principle of maximum a posterior and weighted exponent.Due to the absence of estimated values from the original form,the EKF and SVSF are modified based on the strategy of a one-step smoothing method.This process allows the system to have the smoothed parameter that can be utilized to proceed with the offline-derivation process.Afterward,the suboptimal values of all unknown parameters can be calculated.However,due to the presence of a multistep smoothing term,the adaptive process needs to be simplified.Therefore,the inaccuracy might occur because of a non-positive definite matrix to the covariances corresponding to either process or measurement.For this reason,the use of the divergence suppression method is also involved.Besides that,both the suboptimal estimate values are also estimated using the unbiased estimation method.In the second case,the conventional EKF and SVSF are assisted by a different approach to the previous one,named Maximum A Likelihood Estimator and Expectation-Maximization.In this design,the adaptive forms are generated by assuming that the updated covariance form of EKF and SVSF are the same.However,the mathematical derivation chokes temporarily due to the presence of estimated values which is unavailable from the original formulation of the EKF or SVSF.Therefore,aiming to cover this lack of estimate values,the EKF and SVSF are modified based on a one-step smoothing method.Furthermore,to prevent the divergence caused by covariances' complexity,the unbiased estimation and innovation covariance estimation is involved.Hence,the proposed methods can recursively update the noise statistic under time integration.All the adjusted parameters based on the previous calculation make the filtering learn and improve without changing their characteristics.Furthermore,the proposed methods are applied to solve the SLAM problem of a wheeled mobile robot.Henceforth,both are named as AEKF-SLAM and ASVSF-SLAM algorithm.The proposed method's verification and validation are conducted in two different cases,the synthetic-based simulation and real-experiment.The synthetic-based simulation considers that the robot moves from the initial position to the goal position by executing wheel rotation ticks per second.The user gives these values,but they are assumed to be always followed by the small additive noise.Sequentially,it measures all distinguishable features by presenting the range and bearing values to the system.Similarly,the measurement values are considered to be noisy.Therefore,the syntheticbased simulation assumes the reference path when the robot is moved using a motion model without any perturbation.The map is supposed to be known by placing some features around the robot path.The motion model and measurement model are designed by adopting the differential steering system and direct point-based observation,respectively.Meanwhile,the Victoria Park dataset recorded by Nebot,2009,at the Australian Centre for Field Robotics is used for the real-experiment.This popular dataset is commonly used to verify the adaption or invention to a 2D online featurebased SLAM algorithm.A path through an area of around 197 m x 93 m is described in this dataset.This sequence consists of 7247 frames,captured over a total period of 26 minutes,along a 4 km trajectory.The data set includes steering and rear-axis wheel sensor readings(odometry)and laser range finder readings(one scan of 360 degrees per second)along with GPS data.A tree detector feature is given for the laser range data along with the dataset.Invariably,they have a wide distance to each other and can be isolated or classified with standard data association techniques.However,spurious data is found in some instances and must be deleted.All of the tests are conducted on a 2.3GHz Dual-Core Intel Core i5,8 GB 2133 MHz LPDDR3.The purpose of this experiment is to evaluate our approach's consistency and to examine the computational complexity.According to these simulations,all algorithms' accuracy and consistency are analyzed/compared in terms of average RMSE and NEES under the Monte Carlo Simulation.The comparison shows that the proposed method,ASVSF-SLAM algorithm,is better than the conventional method.