Research On Key Technologies Of Robot Navigation Based On Neural Network Fusion Vision SLAM

SLAM is a key component in the field of robotics research,as well as the basis and prerequisite for autonomous mobile robots to complete various tasks in a satellite denial environment.As a low-cost solution for autonomous mobile robots to perceive the environment,visual SLAM has developed rapidly with the development of image processing technology.In recent years,the fusion of visual SLAM and neural networks has provided a new way to improve the anti-interference,high adaptability,and reduction of computational complexity of visual SLAM algorithms.Therefore,in-depth study of key technologies such as the combination of visual SLAM and deep learning,the fusion of visual sensors and other sensors in SLAM,has important application and theoretical significance for improving the application of visual SLAM in practice.Based on the research of visual SLAM algorithms such as monocular positioning,monocular pose estimation,and visual odometry,this thesis further carries out the monocular vision positioning algorithm of the fusion of vision and neural network,and the visual position of the fusion of vision and semantic segmentation.Pose estimation algorithm,based on neural network vision and inertial fusion mileage calculation method,and based on NVIDIA Jetson TX2 module and mobile phone platform for SLAM map construction related experiments,the specific content mainly includes the following aspects:Firstly,the research of monocular vision positioning algorithm based on neural network is carried out.Collect image data in the sports scene,and recover the scene 3D point cloud through the motion recovery structure,and realize the location and annotation of the scene image through the 3D point cloud;input the image and the annotation data into the positioning neural network to obtain the trained scene model;Finally,input the image to be queried into the model,and output the predicted positioning.The positioning error analysis is carried out on the public data set and the data set collected in the experimental environment.The experimental results show that the positioning speed and accuracy of monocular vision can be effectively improved based on the neural network model.Secondly,the study of monocular vision pose estimation algorithm based on semantic segmentation is carried out.In the environment where there are moving objects,perform semantic segmentation on the original image to obtain the position data of people or animals in the scene;after extracting the feature points,remove the distributed feature points of the dynamic part,and remove the feature points of the remaining area Used for monocular visual pose estimation.The semantic segmentation optimization monocular pose estimation algorithm test was carried out on the TUM public data set.The results show that compared with the monocular pose estimation method in the traditional SLAM algorithm,the semantic segmentation optimization can significantly reduce the error of monocular vision pose estimation.Improve the accuracy and robustness of the monocular visual pose estimation algorithm in dynamic scenes.Then,the study of visual-inertial fusion mileage calculation method based on optical flow neural network was carried out.The sequence image and IMU data are synchronized in time.The optical flow neural network Flow Net Simple method is used for image feature extraction.After convolution,the feature map is directly output as the input of the cyclic neural network,and the output of the optical flow neural network is expanded and the IMU The six-dimensional data and the seven-dimensional pose data of the previous moment are also used as the input of the recurrent neural network.Finally,the estimated robot position and posture are output,and the pose estimation is continued to be input to the recurrent neural network as the next state Enter the amount.The accuracy of the odometer is analyzed on the public data set and the data set collected in the experimental environment.The experimental results show that the optical flow neural network can improve the adaptability and anti-interference ability of the visual-inertial fusion odometer in fast motion and harsh environments.Finally,based on the robot operating system carried by the TX2 development board hardware platform,the SLAM mapping experiment system was designed,and the monocular vision mapping experiment,the monocular camera and the depth camera fusion dense mapping experiment,and the laser two-dimensional and three-dimensional The threedimensional mapping of monocular and laser fusion,respectively constructed sparse point cloud map,dense point cloud map,two-dimensional grid map and other SLAM maps with navigation,positioning,and obstacle avoidance functions.Based on the Android mobile phone platform,the feature point method visual SLAM mapping experiment and the direct method visual SLAM mapping experiment were carried out to verify the performance of camera pose estimation and map construction on the mobile phone.Based on the research needs of visual SLAM and neural network fusion technology,this thesis conducts in-depth research on monocular visual positioning,posture estimation,and fusion visual odometry based on extensive research on the research results of SLAM fusion algorithms at home and abroad.