Simultaneous Localization And Mapping For Mobile Robot Navigation With Stereo Camera

With the development of technology,mobile robots have become an important research topic.To perform the task of autonomous navigation,mobile robots need to have the functions such as mapping,positioning,and path planning.Simultaneous localization and mapping(SLAM)is the process of creating a map and determining its own position at the same time in an unknown environment.SLAM plays a vital role during the whole process of autonomous navigation.According to the difference of used sensors,there are two categories of SLAM,i.e.,visual SLAM and lidar SLAM.Stereo camera sensors have lower cost and richer image information than lidar,so visual SLAM have become a hot point.However,vision sensors are sensitive to changes of light.When the light changes,feature points of the image are unstable,which causes the fails in the visual SLAM.In order to improve the robustness of the visual SLAM system,this thesis has carried out research on the simultaneous localization and mapping for mobile robot navigation with stereo camera.The main contents of this thesis are as follows:(1)An improved ORB image feature extraction and matching algorithm is proposed.Features extracted by traditional algorithms are unstable.This thesis improves it by introducing methods such as non-maximum suppression and automatic threshold setting.The results show that the improved algorithm improves the image feature extraction and matching speed,and the extracted features are more stable and more robust than traditional algorithms.(2)A visual odometry method based on SuperPoint neural network self-supervised learning features is proposed.The visual odometry method based on artificial design features is relatively inefficient and the estimated pose is possibly inaccurate.This thesis introduces the self-supervised feature points and descriptors of the SuperPoint network,and then selects stable feature points to estimate the position.The results show that the number of features extracted by the self-supervised learning feature method is more stable,the estimated position is more accurate,and the cumulative error generated is smaller.(3)A loop closure detection algorithm based on ResNet neural network is proposed.The visual odometry is prone to accumulative errors and the accuracy of traditional algorithms is low.ResNet neural network is used to extract the image features.The extracted features are more robust than traditional methods,and then use ZCA(Zero-phase Components Analysis)whitening to reduce the dimension of image feature vectors.Finally,loop closure detection is performed through a defined similarity matrix.The results show that the proposed algorithm has higher accuracy and solves the problems of low accuracy and low robustness of traditional algorithms.The above algorithm was tested on a mobile robot platform,and a visual SLAM test was performed on public data sets and indoor environments,which proved the effectiveness of the algorithm and improved the adaptability of visual SLAM.