Research On Monocular Visual-inertial Odometry Based SLAM System

Visual inertial odometry has become an important research part of autonomous navigation and positioning of mobile robots in recent years.Visual information and inertial information are highly complementary.Inertial information can be positioned under the condition of rapid camera movement or low environmental texture,while visual information can assist inertial information to correct the influence of its own drift.Therefore,the integration of the camera and inertial components greatly improves the robustness of odometry positioning.However,because visual information and inertial information each have their own limitations,and they complement each other when optimizing posture,accuracy and robustness need to be further improved when positioning in challenging environments such as long-term low light conditions.To this end,this paper designs two kinds systems of visual inertial odometry,and combines the overall framework of real-time localization and mapping(SLAM)to verify the effectiveness of the method based on public datasets and actual indoor scenes.Firstly,this paper introduces the necessary basic theoretical knowledge such as camera model and rigid body motion in three-dimensional space.For IMU's measurement data processing,we uniformly adopt pre-integration algorithm to avoid repeated integration caused by the change of initial value and increase calculation consumption.In this part,we fully consider the influence of Gaussian white noise and drift on IMU's pre-integration,and construct IMU residual model for subsequent pose optimization.In this paper,an edge-guided visual inertial odometry is first proposed.The algorithm uses edges to guide feature points extraction to screen features with higher robustness,and discards some feature points with low discrimination that are easy to cause mismatching to improve feature matching rate.And in the construction of the nonlinear optimization model,iterative optimization is carried out by combining the reprojection error based on the edge guide feature points,the photometric error term based on the edge sampling points and IMU residual equation.Through the above improvement,it can be ensured that the frame is not lost due to insufficient feature point extraction,and the positioning accuracy is improved to a certain extent.In view of the positioning situation in extremely low light environment,this paper proposes a visual inertial odometry based on GAN(Generic Adversarial Network)network,and introduces CycleGAN(Cycle Generic Adversarial Network)network to realize the conversion from low light image to normal light image without one-to-one correspondence of training data.At the same time,in order to make the images processed by the network better adapt to the requirements of continuous positioning and ensure the continuity of image sequences in time and space,this paper introduces the loss function of time sequence consistency based on CycleGAN network.In addition,in pose estimation,this paper innovatively combines the most effective CNN network,FlowNet 2.0,and the traditional visual inertial odometry framework to improve the robustness and accuracy of the algorithm in challenging environments.The visual inertial odometry pose estimation process only uses the feature correlation of local regions,and cumulative errors are inevitable.However,the back-end optimization and loopback detection process can minimize cumulative errors and make the system run for a longer time with higher accuracy.Therefore,in the experimental stage,this paper constructs a complete SLAM system framework by combining the proposed visual inertial mileage calculation method with the back-end optimization and loopback detection process.For SLAM system based on edge-guided visual inertial odometry,experiments on feature matching accuracy and location are carried out based on EuRoc dataset.The results show that the feature matching accuracy is improved by about 3.2% on average,and the accuracy and robustness of pose estimation are improved compared with the most advanced algorithms at present.For SLAM system based on GAN network visual inertial odometry,we used EuRoc public dataset and laboratory mobile robot to test.In order to demonstrate the applicability of the algorithm to challenging environments such as low light,we specially tested the four most difficult image sequences in EuRoc dataset.However,in the actual scene test of mobile robots,the test is carried out in the teaching building hall and the environment keeps extremely low illumination.Compared with other excellent visual inertial SLAM frameworks,the accuracy of our improved algorithm positioning and its effectiveness and reliability in low-light environment are verified.