Research On SLAM Algorithm Based On Binocular Vision And IMU

With the development and popularization of autonomous driving technology,the hardware equipments are also developing towards miniaturization and low cost.In the field of SLAM,information-rich and low-cost cameras have received more and more attention.And multisensor fusion is replacing single camera as the mainstream solution due to the problems pure vision solutions existing,such as relying heavily on environmental textures,being sensitive to light changes and difficulty in coping with scenes that change too fast.Among all kinds of sensors,the inertial measurement unit is extremely highly accurate in a short time,so it is also has been widely used.In addition,due to the lack of loops when building large-scale outdoor maps,it will inevitably cause accumulation of errors.Based on the above questions,this article has done an in-depth study,and we’ll explain them in three parts.(1)We propose a method of stereo mismatch elimination based on adaptive threshold.According to the stereo camera imaging model,the disparity of a point is inversely proportional to its depth.So that we can eliminate mismatched points by limiting the disparity.We propose a disparity threshold that adaptively changes based on the pre-order frames.The threshold can be quickly adjusted while the scene changing.The threshold can eliminate mismatches to the greatest extent,thereby improving the accuracy of subsequent pose estimation.(2)We propose a pose and trajectory optimization method based on OpenStreetMap.Aiming at the accumulation of errors in large-scale loop-free mapping scenarios,we combine the OpenStreetMap with pose estimation and take OSM as prior information.With performing a series of trajectory matching,pose correction and optimization to limit the trajectory in the passable area of the OSM,the accumulation of errors can be effectively reduced.(3)According to the combination method of monocular and IMU,this paper proposes a method of fusing stereo and IMU.We propose a method for joint initialization of stereo vision and IMU,which is made up of three parts: visual initialization,IMU initialization,and joint initialization.And we propose a joint optimization method of stereo vision and IMU,which implements the tight coupling optimization of vision and IMU by unifying the visual error and IMU error into a single framework.We have implemented and experimentally verified the above three parts respectively,and ORB-SLAM2 is used as the visual front end in the latter two parts.The above experiments were carried out on public datasets such as KITTI and Euroc and the results are compared with other open source algorithms’.The experimental results indicate the effectiveness of these methods.