On Pose Estimation Problem Of Wheeled Robot In Complex Environment

With the development of artificial intelligence and robotics,various robots have been applied successfully in scientific research,industry manufacturing,space explo-ration,post-disaster relief,intelligent transportation,and agricultural production,to as-sist or even replace human to finish the simple or dangerous work.Because of the advantages of simple structure,stable performance and low cost,wheeled robots have become the most widely-used robot.Autonomous mobility is the most fundamental function of a wheeled robot.In order to realize accurate position control,a prerequisite is to mainain the accurate per-ception of the robot's pose.However,for a wheeled robot working in complex environ-ment,to implement the pose estimation with low cost,high accuracy,low time-delay,high stability and reliability is still a hot issue.Because there exist a mass of unpre-dictable interference in the environment,which gives rise to anomalies in sensors,it brings a lot of challenges to the robot pose estimation.This dissertation starts from realities and fully considers the issues that may arise in complex environments,such as terrain variation,tire-ground slip,imprecise kinematic model,unknown or slowly-changing statistical characteristics of sensor noises,sensor anomalies,etc.,by means of machine learning,state estimation,image processing,in-terval analysis and some other technologies,to propose a series of solutions for pose estimation of wheeled robots under different scenarios.The main research contents and innovations are as follows:Considering the potential issues rendered by slip,the traditional kinematic model is improved.The steering resistance coefficient is introduced to describe the wheel-ground slip effect,and a more accurate time-discrete form of the kinematic model is obtained based on the motion decomposition.Based on the improved kinematics model,the accuracy of dead reckoning method can be increased.Considering the problem that the traditional odometry is easy to diverge,an adap-tive odometry based on terrain classification is proposed.A terrain classification frame-work combining a simple compact features,a weak classifier and a Bayesian filter is proposed.It can ensure the accuracy of terrain classification while reducing the compu-tational complexity.The robot observes its current terrain in real time and obtains the corresponding steering resistance coefficient of the terrain.It can adapt to the terrain variation,and thus retard the divergence of odometry.Considering the unreliable characteristics of the compass and beacon positioning system,a pose estimation method under controlled environment is proposed.Using the prearranged auxiliary strips and floor vision,the robot can estimate its current pose ac-curately in real time.Since the compass and beacon positioning system are avoided,the reliability of the method is improved dramatically.This method is suitable for industrial scenes that require long-term operation.Considering the issues caused by terrain variation,unknown and slowly-changing noise statistics,sensor anomalies in complex environments,a pose estimation method under uncontrolled environment is proposed.We designs data fusion algorithms for gyroscope-compass combination,encoder-compass combination,gyroscopes-encoders-compass combinations,and heading angle estimators-beacon positioning systems com-bination,to achieve accurate and reliable robot pose estimation in uncontrolled envi-ronments.