Research On Mobile Robot Pose Estimation Based On Deep Learning And Multi-sensor Fusion

In this paper,the pose estimation of mobile robot is studied.According to the requirements of high robustness and high accuracy for pose estimation in complex scenes,multi-sensor information fusion and depth learning methods are adopted respectively.According to the two methods proposed,the framework of pose estimation of mobile robot is studied and designed,including the hardware structure of serial port driver,mathematical operation library,multi view geometry and motion model,the software environment of operating system,deep learning framework,cloud server,and the slam system of motion perception,parallel computing and functional algorithm.Firstly,the traditional visual slam is studied.In the part of visual odometer,feature point method and direct method are used to realize,sparse pose trajectory and scene map are established by using the generated visual features,and back-end optimization is used to adjust pose and map based on filtering and nonlinear optimization respectively,so as to gradually reduce the uncertainty of pose estimation.In order to build a globally consistent pose trajectory,the loop detection based on the word bag model is added.The experimental results show that the matching error is reduced to 0.6m after adding loop detection,while the RMS value of absolute trajectory error using nonlinear optimization is significantly smaller than that based on filtering over time.Secondly,the paper studies the vio algorithm of IMU and vision information fusion.Firstly,the residual model of vision ins fusion based on nonlinear optimization is constructed,and its Jacobian matrix is explained in detail.In order to avoid repeated motion integration of IMU data during attitude updating,the relative motion integration of IMU data between keyframes is performed by using the IMU pre integration method.IMU data is regarded as motion information,and visual observation information is used for data association and motion information correction.In the experimental part,the direct method and the feature point method are used to realize the vio.The results show that the fusion algorithm can still run correctly with obvious motion blur and illumination change,which shows the robustness and high precision of the fusion algorithm.In addition,the paper proposes a strategy to improve the fusion efficiency of vision and IMU,and adopts the idea of a graph model.In order to improve the calculation efficiency after data fusion,the older observation in the state quantity is deleted selectively by using the method of marginalization,but part of its information is retained as the prior of the next observation.This strategy is used in the slam system,and the experimental results verify the feasibility of this method.This paper proposes a compound trestle self encoder which is suitable for loopback detection.Using the multi-layer trestle self encoder,an object classification model is established.The similarity between the decoded image features and the features captured by the camera is calculated,and the similarity comparison threshold is designed.The experimental results show the feasibility of this method.Compared with the commonly used word bag model,it is based on deep learning The accuracy of loop detection is improved significantly.