Visual Inertial Positioning Algorithm Based On Farmland Scen

Visual-Inertial Odometry(VIO)is a localization algorithm that uses monocular camera images and inertial sensor information.The VIO algorithm is low cost,lightweight and high quality.Therefore,it has been widely used in positioning and navigation algorithms for unmanned delivery vehicles,UAVs,autonomous robots,and portable devices in both urban and indoor scenarios.While monocular visual inertial localization algorithms(VIOs)have performed well in urban and indoor environments,their accuracy and robustness in localizing agricultural automation equipment is significantly reduced in specific environments such as farmland.This is due to factors such as unstructured scenes with unstable features,changing light conditions,and rugged terrain.To address these challenges,there is a need to improve the accuracy and robustness of the monocular VIO algorithm for localization in farming scenes.Therefore,this paper improves the existing state-of-the-art monocular VIO algorithm VINS-mono by focusing on two modules: the vision processing front-end module and the fault detection and recovery module.(1)In the vision processing front-end module,an image key frame selection algorithm based on vertical motion smoothness verification is proposed in this paper.This algorithm can be used to remove blurred image frames and image frames with unfocused field of view,thus reducing the rapid loss of image feature points during tracking and thus improving the accuracy of the localization algorithm.(2)In the fault detection and recovery module,a trajectory tracking recovery method based on the assumption of local road flatness is proposed in this paper.The method can quickly recover the motion trajectory tracking process interrupted by faults,effectively alleviating the faults caused by the drift of the trajectory direction and the failure of the visual sensor due to light changes during the localization process,thus improving the robustness of the localization algorithm.(3)To increase the practicality of the monocular VIO localization algorithm,this paper also integrates a crop row sensing module in the localization algorithm,which is used to sense the location of crop rows in the farm environment and generate a map with crop row distribution information.The environment map can be used for high-level decision making and help the automated equipment to better complete farming operations.However,the mainstream VIO algorithm can only be used to provide bit position estimation of the mobile platform and cannot sense the structural information of the environment.Experiments on the Rosario dataset collected by farmland weeding robots demonstrate that the algorithm proposed in this paper outperforms VINS-mono in terms of localization accuracy and robustness,with a 69% reduction in absolute trajectory error.In addition,this paper adds a crop row sensing module,which is found to identify and locate crop rows in farmland by comparing the visualization results,which effectively improves the practicality of the monocular VIO localization algorithm.Therefore,based on the experimental results and visualization comparisons,the algorithm proposed in this paper has better applicability in farming operations and shows higher localization accuracy and robustness.