| Visual SLAM(Simultaneous Localization And Mapping)is the main technology to realize 3D reconstruction And Localization in a specific environment based on camera perception environment.In recent years,with the mobile robot for areas and business needs continuous amplification,visual SLAM because of its good positioning performance in a specific scenario become a hot research direction in the field of mobile robots,but then on the stability and precision of robot localization system have higher requirements,a series of problems still exist in the oriented to complex scenarios,The performance of existing monocular SLAM methods is easily affected by environmental factors,such as abnormal image exposure caused by complex lighting,wrong matching points caused by repeated textures and dynamic objects,and degradation caused by a single sensor.In order to improve the localization performance of visual SLAM in complex environment,aiming at the problems and characteristics of complex environment,this paper mainly studies the SLAM algorithm based on binocular and inertial navigation.The main contents of this paper are as follows:Firstly,aiming at the light sensitivity problem of visual SLAM in complex lighting scenes,a binocular visual odometry integrated with online photometric calibration is proposed.Using the feature point constraint relationship detected by the visual front end,the photometric imaging parameters are continuously measured on continuous frame images.The estimation and update are performed to realize the photometric correction of the input image,so that the front end of the SLAM system is no longer disturbed by the image grayscale changes caused by camera imaging,which is beneficial to improve the illumination robustness of the visual SLAM system.Secondly,in view of the wrong matching problem caused by dynamic objects and repeated textures in complex scenes,an outlier elimination method based on inertial navigation uncertainty model is proposed.According to the inertial navigation kinematics and pre-integration model,the projection of spatial points is established.Position prediction,furthermore,the selected points are judged and eliminated by chi-square detection.Thirdly,for the back-end optimization problem of multi-sensor fusion,a lightweight graph optimization method combining visual residuals and IMU residuals is proposed,which is optimized only for keyframes with strong common-view relationship to establish constraints to ensure that the two sensors Efficiency and precision of joint optimization.Finally,in the public data sets and real scene to photometric correction of this method,the positioning ability experiment test,and the experimental results show that the algorithm in complex and dynamic objects,repetitive texture and light remained higher image fuzzy scenarios,such as the stability,with other mainstream visual SLAM method of comprehensive experiment and comparison,The algorithm also shows some advantages in the performance evaluation of real-time and positioning accuracy in some data sets. |
