| With the development of computer technology and intelligent manufacturing industry,the field of intelligent mobile robots has attracted more and more attention,and Simultaneous Localization and Mapping(SLAM)is the basis for intelligent mobile robots to work in unknown environments.SLAM mainly helps the robot to perceive its position in 3D space and the surrounding environment structure through sensors,so as to implement autonomous positioning,map construction and path planning.With the application of technology entering a new stage,the scenes faced by SLAM are more complex and changeable,which places new requirements on the original technology.The traditional feature extraction algorithm used by the traditional SLAM system relies on manual design,which is effective and practical in some specific scenes.However,when faced with scenes such as missing textures,their performance is not stable and may even fail to run.In addition,the visual SLAM system is difficult to meet the requirements of precise positioning in the case of image blurring and too few overlapping areas of two frames caused by too fast camera motion.Therefore,this paper conducts research on the multi-task feature extraction network,fusion of visual inertial sensors,and 3D reconstruction technology,and proposes a visual inertial SLAM based on multi-task feature extraction network.The research content of this paper includes the following parts:Aiming at the problem that the traditional feature extraction algorithm used by the traditional SLAM system is difficult to deal with texture-less regions,a multi-task feature extraction network with higher recognition rate and matching rate is proposed to replace the traditional feature extraction algorithm for feature detection.The features generated by the network keep the same descriptor format as the ORB features and have good portability.At the same time,it basically meets the real-time requirements of SLAM applications.For the visual odometry transplanted with multi-task feature extraction network,experiments are carried out in the public datasets and real environments.The experimental results show that the network is portable and effective in improving the pose estimation accuracy of SLAM system.Aiming at the problems of image blurring and too few overlapping areas of two frames caused by too fast camera motion,through the fusion of visual sensor and inertial sensor,a visual inertial SLAM based on multi-task feature extraction network is proposed.By pre-integrating the IMU data,a loosely-coupled method to determine the initial value is adopted in the front end of the system,and the visual coordinate system is aligned with the inertial coordinate system.In order to realize the tight coupling of visual inertia information at the back end of the system,a combined energy function that integrates stereo photometric error energy function and IMU error energy function is proposed.By minimizing the combined energy function,a fixed number of keyframe pose and IMU error are optimized in the sliding window.In order to ensure the global consistency of 3D model in the process of 3D reconstruction,and make the reconstructed model cover the 3D environment as densely as possible,a 3D reconstruction method based on surfel model is proposed.Through the surfel model,The proposed method obtains the individual information of the point cloud,divides the point cloud into the newly reconstructed point cloud and the previously reconstructed point cloud according to the time node,aligns the two point cloud regions to implement the point cloud fusion,and uses the combination of four cost functions to implement the optimization of the point cloud.In this paper,the visual inertial SLAM based on multi-task feature extraction network is deeply studied.According to the comparative experiments of different scenes,the visual inertial SLAM based on multi-task feature extraction network proposed in this paper can effectively overcome the problems of camera jitter and too fast scene change,and has more advantages than traditional methods in the overall accuracy and stability of pose estimation.At the same time,the system proposed in this paper can better reduce the dislocation and local distortion of 3D models,so as to improve the overall accuracy of 3D reconstruction.Therefore,the method proposed in this paper can be widely used in the fields of intelligent mobile robot,augmented reality and virtual reality. |
