Research On Simultaneous Localization And Mapping For Robot Based On Hybrid Features

With the development of artificial intelligence industry,it is necessary for the robot having much more intelligence.The technology of SLAM(Simultaneous Localization and Mapping)is widely used in service robots,automatic driving,VR(Virtual Reality)and other fields due to its rich information and low cost.At present,visual SLAM mainly includes feature point method and direct method.The feature point method mainly uses the point features in the image to estimate the pose of robot,it may affect the algorithm effect when some environments contain fewer point features.In this thesis,the simultaneous localization and mapping algorithm for robots based on hybrid feature model is proposed in order to solve the problem.Firstly,an omnidirectional mobile platform is built to collect image information,and an accelerated matching model based on environment dictionary is constructed.Secondly,the hybrid feature model proposed in this thesis is used to estimate the pose,which makes up for the problem that the feature point method is less effective in the environment with insufficient point features.Finally,the optimization strategy based on the maximum common-view weight frame is used to optimize the pose and complete the map construction.The research of the thesis mainly includes the following aspects.(1)Construction of binary environment dictionary and accelerated matching model.For the feature point SLAM algorithm,the text type dictionary is usually loaded with a long time and the scene is less targeted.This thesis builds a binary environment dictionary and feature acceleration matching model.It mainly include three steps to get the environment dictionary.Firstly,the feature points of all the images in the set and the descriptors for each feature points are obtained.Secondly,the image point features are clustered by using bag of words library.Thirdly,the dictionary in the text file format is converted into binary type file by using boost libraries.The accelerated matching model reduces the feature matching range.Comparing with the text type dictionary,the binary environment dictionary can get a better estimation result.(2)Pose estimation algorithm based on hybrid feature modelPoint feature is widely used in position and pose estimation in visual SLAM,but it has certain requirements for scene.In the case of insufficient or missing point features,the pose estimation effect is poor.This thesis proposes a pose estimation method based on hybrid features.The proposed algorithm first extracts the point-line features in the image and uses the improved random sample consensus to eliminate mismatched features.Then,the efficient perspective-n-point method is used to estimate the pose of robot.In addition,this thesis proposes a method based on the line segment estimation in a single image.The fusion of the two estimation results improves the robustness of the algorithm.(3)Global optimization based on maximum common view weight frame and map buildingAiming at the problem that the calculation is large for the back-end optimization and the point cloud map occupies large memory,this thesis proposes a global optimization method based on the maximum common-view weight frame.After the keyframes are obtained,the number of common-view map points among the key frames are counted as the common view weight.The keyframes involved in the optimization select by setting the common view weight threshold.According to the result of optimization,the point cloud map is converted into octree map and finally the globally consistent map is generated.The thesis proposed the method of simultaneous localization and mapping for robot based on hybrid features algorithm.The comparisons experiments have been done based on fr1/room and fr2/desk in the RGBD datasets and the experimental results verified the effectiveness of the proposed algorithm.Comparing with the ORBSLAM algorithm only used point feature,it can be seen that the square root error of the camera trajectory decreased 2.27 cm,4.87 cm respectively.The map construction in the actual environment experiment was carried out by using the self-constructed omnidirectional mobile platform,the experimental results show the effectiveness of the simultaneous localization and mapping for robot based on hybrid features algorithm.