| Simultaneous Location And Mapping(SLAM)refers to a mobile robot using its own carried vision sensors to collect environmental information without acquiring current environmental information,and build an environmental perception model during motion and estimate its own motion information at the same time.The study of visual SLAM environment perception technology for mobile robots is crucial to improve mobile robots to achieve high intelligence and high decision making in complex scenes,but pure visual SLAM algorithms cannot accurately perceive environmental information to achieve accurate localization in fast motion or weakly textured environments by relying only on a single visual sensor.As an aim at enabling visual SLAM set of computer instructions to successfully deal with the complex and surrounding conditions,multisensor fusion localization technology has become a research focus,among which the positioning and navigation technology of visual inertial combination is popular,but how to make real the multi-sensor information joining together of different frequencies is still a problem to be solved very badly.This paper focuses on the in-depth research and exploring things of the SLAM optimization method for mobile robots with vision and the force of something movingl fusion,and focuses on the problems of real-time and robustness of the visual-inertial optimization method in fast motion and weak texture scenes.To deal with problems of poor real-time performance and low robustness of traditional visual SLAM in the feature extraction segment,a combination of improved ORB feature point extraction and improved optical flow matching is used.In the feature extraction process,a multilayer Gaussian image pyramid is constructed,and each layer is meshed and the feature points are homogenized using a quadratic tree to reduce the inter-frame pose solving errors caused by feature redundancy.In the feature matching process,we use the combination of multi-scale image pyramid and iterative optical flow to reduce the resolution of image pixel points by region segmentation,and reasonably select the pixel points in each segmentation region to calculate the motion change of a pixel point in different resolution images,so as to obtain more detailed and accurate pixel coordinates of pixel points and enhance the robustness of the system.In order to solve the problems such as the data association of feature point matching errors in weak texture scenes,we design an error tracking point rejection mechanism based on the reprojection error and RANSAC principle to exclude the error tracking points and construct a local trajectory map by solving the positional changes between successive image frames of the camera through the reprojection construction to the polar geometric constraint method.To address the problems of poor localization and large errors of pure visual SLAM in the face of fast motion,drastic scene illumination changes and texture-deficient scenes,visual sensors are fused with inertial sensors to make up for the shortcomings of visual SLAM in fast motion and weak texture scenes by using the complementary nature of both.In order to solve the problem of effective fusion of visual inertial data,this paper proposes a nonlinear tight-coupling optimization of sliding window based on covisual constraint,using the co-visual relationship between key frames to grade key frames,keeping the co-visual strength between key frames as a constraint to optimize only strong co-visual key frames in the sliding window,and using DogLeg calculation method to reduce the optimization computation and speed up the computation,eliminate the local cumulative error,and improve the back-end optimization accuracy.The back-end optimization accuracy is improved.To verify the effectiveness of the proposed algorithm,the multi-scale optical flow fusion feature-based visual SLAM algorithm and the optimal estimation of multi-bit pose information based on key frames are tested in the publicly available KITTI and EuRoC datasets,respectively,and compared with the mainstream ORB-SLAM,ORB-SLAM2 and VINSmono algorithms to verify the performance of the proposed algorithm in fast motion and weak texture The algorithm is compared with the mainstream ORB-SLAM,ORB-SLAM2 and VINS-mono algorithms to verify the real-time and robustness of the proposed algorithm in fast motion and weak texture scenes.The algorithm is used on the built hardware platform to demonstrate the adaptability of the algorithm and to provide a reference for the research application of monocular/inertial SLAM in indoor environment. |
