| Low light environment can cause images to lack clear textures and environmental details,severely affecting the extraction of image features by visual SLAM algorithms,which results in the inability to complete localization and mapping tasks under low light environment.Direct methods for SLAM algorithms have the advantage of being applicable to weak texture scenes and can establish semi-dense point cloud maps.Therefore,this thesis proposes an improved algorithm based on direct visual-inertial odometry to improve localization accuracy and mapping effectiveness in low light environment.The main work of this thesis includes:(1)Establishing mathematical models for camera optical imaging,pinhole imaging,and camera distortion,analyzing the motion model of the inertial measurement unit(IMU),introducing the theory of IMU preintegration,and defining the coordinate system and transformation relationship of the unmanned vehicle localization system.(2)To address the problem of insufficient image information in low-light conditions for visual SLAM algorithms,a deep learning-based image enhancement algorithm is employed using the LED-NET network to enhance the low-light images and remove blur.To solve the problem of non-repeatability of gradient points extracted by direct methods,which cannot be used to construct a dictionary for loop detection,ORB feature points are extracted for feature description,and the quadtree algorithm is used to achieve a uniform distribution of feature points in space.(3)To address the problem of accumulated errors in visual-inertial localization algorithms during long-term operation,a loop closure detection module is designed based on an online bag-of-words model.To avoid introducing false positive candidate loop frames,co-visibility constraints and epipolar geometry constraints are used for candidate loop frame screening and validation.Based on this,an error function is established to optimize the pose transformation between the current frame and the correct loop frame,and then the global pose graph is optimized and map points are updated based on the detected loop constraints and adjacent keyframe constraints.(4)Experimental verification of the proposed improved algorithm is conducted.The TUM-VIO dataset and the 4Season dataset are used to verify the improvement of the proposed algorithm in low light environment for localization accuracy.The industrial and commercial park scenes in the 4Season dataset are used to verify the effectiveness of the loop closure detection algorithm proposed in this thesis.Finally,real vehicle experiments are conducted in the campus environment to verify the effectiveness of the proposed improved algorithm in real low light environment.The experimental results of the dataset and the real vehicle experiments show that the proposed improved monocular visualinertial localization algorithm can provide accurate localization services in low light environment and has good practical value. |
