Research On Visual Homing Algorithm Of Intelligent Mobile Robot

Robot autonomous navigation is a widely discussed topic in the field of intelligent robots,and the visual sensors-based robot visual navigation technology is one of the main research directions.Depending on the way of processing,visual navigation can usually be divided into two categories.One is the quantitative visual navigation method,which is represented by Simultaneous Localization and Mapping(SLAM)technology.Since this type of method usually requires the robot to store environmental information and positioning information,and needs to continuously update the information during the moving process,such visual navigation methods often have high computational complexity and are also prone to cumulative errors The other one is the qualitative visual navigation method,which is represented by visual homing technology.This type of method only requires the robot to extract the visual landmarks in the current environment,and calculate the homing direction according to the difference of the landmarks' orientations at the robot's different locations,so it has the advantages in terms of simple model and low calculation amount.In this paper,visual homing is studied in detail.In this paper,the visual homing system is designed,and a set of homing performance evaluation criteria is proposed,which provided a theoretical basis for the subsequent experiments.In order to test the performance of the visual homing algorithms in practical applications,an intelligent mobile robot platform and a catadioptric panoramic vision imaging system were built,and a specific scheme for the mobile robot to perform the homing task was specified.In addition,in order to quantitatively compare the performance of different homing algorithms,this paper also summarizes and designs a total of seven performance metrics,including homing vector field,average angular error,angular error grid,return ratio,average homeward component,total distance error and average trajectory smoothness.Using the above performance evaluation indicators,it is possible to more objectively evaluate the performance of different homing algorithms in different aspects.This paper studies the matching precision and spatial distribution optimization of visual landmarks in the field of visual homing,which effectively improves the homing accuracy of the visual homing algorithm.Visual landmark is the only effective input of visual homing algorithm.However,the visual homing algorithm with natural landmarks as effective input often needs to follow the principle of uniform distribution of landmarks.When the detected natural landmarks are not accurate or unevenly distributed,the accuracy of the visual homing algorithm will be greatly affected.In this paper,SIFT features are used as natural landmarks.According to SIFT feature scale information and panoramic vision imaging principles,three landmark accuracy and distribution optimization strategies are proposed,including mismatch elimination strategy,landmark contribution evaluation strategy and SIFT multi-level matching strategy.In this paper,the Average Landmark Vector(ALV)algorithm and the Warping Based on SIFT Features(S-Warping)algorithm are taken as examples to verify the effectiveness of the three landmark accuracy and distribution optimization strategies in different environments.This paper introduces the ALV algorithm and the Homing in Scale Space(HiSS)algorithm in detail,including the specific steps of the two algorithms to achieve homing and their respective advantages,and proposes the existing problems of the above two algorithms.An optimization mechanism called Vector Pre-assigned Mechanism(VPM)is presented in this paper.According to different implementation methods,VPM can be divided into the static VPM and dynamic VPM.The simulated homing experiment and the actual homing experiments show that the ALV and HISS algorithm with VPM have greatly improved the homing accuracyAiming at the problem that most visual homing algorithms have a significant performance degradation in complex dynamic environments,a novel visual homing algorithm called Machine Learning-based Homing(MLBH)algorithm is proposed.As a type of the visual homing algorithm which treats the SIFT features as natural landmarks,MLBH classifies the landmarks according to the scale values of the two features in the same feature matching pair,and uses the Support Vector Machine(SVM)to obtain the decision boundary of the two types of landmarks,thus generating the corresponding homing vector.In this paper,the Average Landmark Vector Based on Sparse Landmarks(SL-ALV)is used as a comparison algorithm to test the homing performance in both static and dynamic environments.According to the simulated homing experiment and the actual homing experiments,the MLBH algorithm shows excellent homing performance and robustness in the dynamic environment.When using different homing performance metrics for quantitative evaluation,the MLBH algorithm is significantly higher than the comparison algorithm.